62 10 3494 https://omeka.ibu.edu.ba/files/original/730719527e5578fd3d43ba30f199617d.pdf 3d79986ea43502148cc1dca1e5cfe0dc PDF Text Text PROCEEDINGS th ______ The 5 International Symposium on Sustainable Development_______ ISSD 2014 COMPARISON OF CODON USAGE IN MITOCHONDRIAL GENOMES OF RHINOLOPHID AND HIPPOSIDERID BATS Semir Dorić1 and Lada Lukić Bilela1,2# 1 Dept. of Biology, Faculty of Science, University of Sarajevo, Zmaja od Bosne 33-35, 71 000 Sarajevo, Bosnia and Herzegovina; semir.doric@gmail.com; llbilela@gmail.com 2 Biospeleological Society in Bosnia and Herzegovina, Avde Jabučice 30, 71 000 Sarajevo, Bosnia and Herzegovina; biospeld2008@gmail.com # Corresponding author Abstract According to current phylogenetic hypotheses, the bats of the families Rhinolophidae and Hipposideridae are sister groups nested within the clade of Pteropodiformes. The Hipposideridae are family of bats commonly known as the Old World leaf-nose bats. While this family has long been considered as a rhinolophid subfamily Hipposiderinae, it is now more generally classified as its own family. The Hipposideridae contain 10 living genera and more than 70 species, mostly in the widespread genus Hipposideros. This study is an attempt to confirm a distinction between these two families by a codon usage comparison of a complete set of mitochondrial protein-coding genes from currently available mitochondrial (mt) genomes of rhinolophid and hiposiderid bats. The INCA 2.1 and GCUA 2.0 were used for the codon usage computing. Measure Independent of Length and Composition (MILC), was used to estimate the codon usage of 13 mt protein-coding genes from five species of genus Rhinolophus and one species of Hipposideros (while only four genes were available from H. larvatus). Large randomly generated sequence sets were used to test for dependence on (i) sequence length, (ii) overall amount of codon bias and (iii) codon bias discrepancy in the sequences. Our findings suggest no significant differences in codon usage bias, among analyzed rhinolophid species, by statistical estimation of absolute frequency values despite the changed MILC values for nd1 and nd3 from Hipposideros armiger. Keywords: MILC, MELP, bats, codon usage, codon frequencies 39 | P a g e �ISSD 2014 th The 5 International Symposium on Sustainable Development_______ PROCEEDINGS 1. Introduction Rhinolophidae split from their sister family, the Hipposideridae, towards the end of the Eocene (Maree and Grant, 1997; McKenna and Bell, 1997; Teeling et al., 2003; Eick et al., 2005). This estimate of divergence is congruent with the fossil data, with fossils of extinct Rhinolophus and Hipposideros species first occurring in middle Eocene deposits (ca. 49–37 MYA; Simmons and Geisler, 1998). The family Rhinolophidae Gray, 1825 consists of a single genus Rhinolophus Lacépéde, 1799. The taxon is exclusively Old World, with at least 77 species (Simmons, 2005) occurring in both temperate and tropical areas throughout the Afrotropical, Australian, Indomalayan, Oceanian and Palaearctic regions (Csorba et al., 2003). A previous classifications imply two subfamilies, the Hipposiderinae and the Rhinolophinae, of the family Rhinolophidae according to Koopman, 1993, 1994; McKenna and Bell, 1997; Simmons and Geisler, 1998; Teeling et al., 2002, while contemporaneously the hipposiderids excluded from this family following Corbet and Hill (1992), Bates and Harrison (1997) and Simmons (2005). Interestingly, Rhynolophus monoceros (used in our study) is often treated as a Taiwanese endemic, and very similar to R. pusillus of the mainland in terms of body size, echolocation call frequency and mitochondrial gene sequences (Li et al. 2006). It is perhaps best treated as a synonym of R. pusillus, especially given that the Taiwanese Hipposideros terasensis considered synonymous with H. armiger of the mainland according to Simmons (2005). Besides, in more recent studies of deep rhinolophid phylogeny based on analysis of cytochrome b and the three nuclear introns: thyrotropin, thyroglobulin and protein-kinase (PRKC1) the hipposiderid bats were used as outgroup (Stoffberg et al., 2010) A multidisciplinary approach: morphometric measurements (Bogdanowicz, 1992), recording and analysis of echolocation signals, karyotypic variation (Koubínová et al, 2010), D-loop sequence analysis) contribute in resolving t1he correct phylogenetic position of the species within these two families (Stoffberg et al, 2010). According to current phylogenetic hypotheses, the bats of the families Rhinolophidae and Hipposideridae are sister groups nested within the clade of Pteropodiformes (Koubínová et al, 2010). Rhinolophidae form a monophyletic group and can be divided into at least two major clades – the predominantly African and the predominantly Oriental clades – based on the current biogeographical distributions of the majority of species within each clade. Morphological (Bogdanowicz, 1992) and cytochrome b (Guillén-Servent et al., 2003) analyses also suggest that the African rhinolophids form a monophyletic clade. The typical metazoan mitochondrial (mt) genome comprises a single circular, double-stranded DNA molecule with a size between 14 and 18 kb that contains a uniform set of 37 genes (Boore, 1999). Mitochondrial genomes are powerful tool in phylogenetic analyses to elucidate the complex relationships among taxa. More rapidly evolving mitochondrial genes may distinguish even closely related species and thus they have been employed in conservation genetic studies (Avise, 1995). Thus, generation of full mtgenome sequences is important for both evolutionary studies and conservation management of endangered species. The aim of this study was a codon usage comparison of mitochondrial genes and identification of possible differences in codon frequency values from rhinolophid and hipposiderid bats as a contribution to phylogeny elucidation between these two families. 2. Materials and methods The nucleotide sequences of 13 (nd1, nd2, nd3, nd4, nd4l, nd5, nd6, cox1, cox2, cytb, cox3, atp6, atp8) mitochondrial protein-coding genes from five species of the family Rhinolophidae and Hipposideros armiger (Hodgson, 1835) as well as four genes (nd1, nd2, cox1, cytb) from H. larvatus (Horsfield, 1823) were obtained from GenBank (NCBI) (Table 1). 40 | P a g e �PROCEEDINGS th ______ The 5 International Symposium on Sustainable Development_______ ISSD 2014 Table 1. Selected species of the families Rhinolophidae and Hipposideridae used for sequence analysis, with a mitochondrial genome accession number in GenBank database. Selected species Accession number Rhinolophus ferrumequinum (Schreber, 1774) NC_020326.1 R. pumilus K. Andersen, 1905 NC_005434.1 R. monoceros K. Andersen, 1905 NC_005433.1 R. formosae (Sanborn, 1939) NC_011304.1 R. luctus (Temminck, 1834) NC_018539.1 Hipposideros armiger (Hodgson, 1835) H. larvatus (Horsfield, 1823) H. larvatus (nd1) H. larvatus (nd2) H. larvatus (cox1) H. larvatus (cytb) NC_018540.1 JX861075.1 DQ888653.1 JQ915493.1 JQ365642.1 EU434949.1 INCA 2.1 (Supek and Vlahoviček, 2006) and GCUA 2.0 (Fuhrmann et. all., 2004) were used for codon usage computing. Measure Independent of Length and Composition (MILC), was used to estimate the codon usage of thirteen mtDNA genes of selected species within genera Rhinolophus and Hipposideros. Large randomly generated sequence sets were used to test for dependence on (i) sequence length, (ii) overall amount of codon bias and (iii) codon bias discrepancy in the sequences. MILC Based Expression Level Predictor (MELP) was used as a measurement to quantitatively predict the levels of selected mitochondrial gene expression. 3. Results and discussion Calculations of MILC and MELP values for thirteen and four protein-coding genes of selected species, six and seven respectively, were carried out in INCA 2.1. (Table 2). Reference MILC values were used as absolute frequencies of codon usage for selected genes (Fo). Estimation of the percentage difference of codon usage bias among selected species followed trough χ2 statistical observations as well as GCUA 2.0. Table 2. MILC and MELP average values for 13 (from five rhinolophid and one hipposiderid bat) (A) as well as four protein coding genes from Hipposideros larvatus (B). Selected species Rhinolophus ferrumequinum R. pumilus R. monoceros R. formosae R. luctus Hipposideros armiger MILC 0.631 0.627 0.635 0.615 0.633 0.666 MELP 1.011 1.009 1.015 1.030 1.021 1.007 A Selected species Rhinolophus ferrumequinum R. pumilus R. monoceros R. formosae R. luctus Hipposideros armiger Hipposideros larvatus MILC 0.516 0.525 0.527 0.511 0.527 0.509 0.522 MELP 1.057 1.050 1.054 1.043 1.043 1.030 1.022 B 41 | P a g e �ISSD 2014 th The 5 International Symposium on Sustainable Development_______ PROCEEDINGS Absolute differences between the theoretical (Ft) and observed (Fo) values of codon frequencies were less than 1%, comparing thirteen and four protein-coding genes from six and seven selected species, respectively (Graph. 1). Graph 1. Absolute difference of Ft and Fo for MILC values for 13 (from five rhinolophid and one hipposiderid bat) (A) and four protein-coding genes from Hipposideros larvatus (B). 42 | P a g e �PROCEEDINGS th ______ The 5 International Symposium on Sustainable Development_______ ISSD 2014 The obtained value of χ2 was 0.00218 with p=0.01 and n=5 (six species) and 0.0096 with p= 0.009 and n=6 (seven species). In both cases p value was <0.5 which indicates no significant difference. However, analysis mediated by GCUA 2.0 revealed important differences in codon usage frequencies. Absolute differences in codon usage frequency between Hipposideros armiger and Rhinolophus ferrumequinum were 8.19% , H. armiger and R. monoceros 8.17%, H. armiger and R. pumilus 9.89%, H. armiger and R. luctus 8.3%. The typical deviation from the universal genetic code observed in mitochondrial genomes of rhinolophid and hipposiderid bats are similar to that already found in other vertebrates, with TGA coding for tryptophan, instead of being a stop codon. Hipposideros armiger prefers TGG (Trp) with the absolute frequency of 0.154 related to five species of genus Rhinolophus with frequencies of 0.05-0.65. Among the terminations codons, UAA was the most preferred by all analyzed species, then AGA while the codon AGG is found neither in mitochondrial genes of five rhinolophid species nor in H. armiger and H. larvatus. AGA and AGG were thought to have become mitochondrial stop codons early in vertebrate evolution (Osawa, et al., 1989). However, at least in humans it has now been shown that AGA and AGG sequences are not recognized as termination codons. UAG codon is rather preferred by H. armiger than rhinolophid bats with a difference of 4.04%. Actually, the UGA Stop-to-Trp is the change is the most frequently occurring reassignment known. Disappearance of UAG would be favored by mutation pressure increasing the AU content. The reason UAG is reassigned less frequently than UGA may be because of the relative difficulty of the required change in the tRNA. In the case of UGA, the existing tRNA-Trp can simply mutate its anticodon (Sengupta et al., 2007). Synonymous codons are not used with equal frequencies, so based on a multivariate analysis of codon usage data from unicellular organisms, Grantham et al. (1980) proposed the genome hypothesis implying some relationship between codon usage and taxonomic distance. A long time ago was noticed a correlation between taxonomic divergence and the similarity of the codon dialect (Ikemura, 1985; Maruyama et al. 1986). 4. Concluding remarks During the last decade, analyses of the mitochondrial genome became a powerful tool to resolve the phylogenetic relationships among the various eukaryotic lineages and to elucidate the early events during evolution of multicellularity. MILC and MELP algorithms have proved to be excellent tools for mitogenomic phylogeny and molecular evolution studies. The codon usage comparison of 13 mt protein-coding genes from five species of genus Rhinolophus and one species of Hipposideros has shown 8.17 to 9.89% differences of codon frequency values using GCUA 2.0. However, our results mediated by INCA 2.1., without any statistically significant difference in codon usage among selected species could be explained by the estimation based on the only one complete mitochondrial protein-coding gene set (H. armiger) and four genes (from H. larvatus) from hipposiderid bats. Furthermore, MILC algorithm has proved to be very sensitive, discriminating genes by their length, alternative stop codons and nucleotide composition. MILC values for nd1 and nd3 genes from hipposiderid bats differ from genes of selected rhinolophid species due to the nucleotide composition. This could indicate possible further differences on complete mtDNA sequences which can help in elucidation of phylogenetic relationships within these two chiropteran families. 43 | P a g e �ISSD 2014 th The 5 International Symposium on Sustainable Development_______ PROCEEDINGS 5. References Agnarsson, I., Zambrana-Torrelio, C., M., Flores-Saldana, N., P., & May-Collado, LJ. (2011). A time-calibrated species-level phylogeny of bats (Chiroptera, Mammalia). PLOS Currents, 10, 13-71. Avise, J.C. (1995). Mitochondrial DNA polymorphism and a connection between genetics and demography of relevance to conservation. Conservation Biology, 9, 686–690. Bogdanowicz, W. (1992). Phenetic relationships among bats of the family Rhinolophidae. Acta Theriologica, 37, 213–240. Boore, J.L. (1999). Animal mitochondrial genomes. Nucleic Acids Research, 27, 1767–1780. Csorba, G., Ujhelyi, P., & Thomas, N. (2003). Horseshoe Bats of the World (Chiroptera: Rhinolophidae). Alana Books, Shropshire. p. 160. Eick, G.N., Jacobs, D.S., & Matthee, C.A. (2005). A nuclear DNA phylogenetic perspective on the evolution of echolocation and historical biogeography of extant bats (Chiroptera). Molecular Biology and Evolution, 22, 1869–1886. Fuhrmann, M., Hausherr, A., Ferbitz, L., Schödl, T., Heitzer, M., & Hegemann, P. (2004). Monitoring dynamic expression of nuclear genes in Chlamydomonas reinhardtii by using a synthetic luciferase reporter gene. Plant Molecular Biologie, 55(6), 869-81. Grantham, R., Gautier, C., Gouy, M., Mercier, R., & Pave, A. (1980). Codon catalog usage and the genome hypothesis. Nucleic Acids Research, 8, 49-62. Guillén-Servent, A., Francis, C.M., & Ricklefs, R.E. (2003). Phylogeny and biogeography of the horseshoe bats. In: Csorba, G., Ujhelyi, P., & Thomas, N. (Eds.), Horseshoe Bats of the World. Alana Books, p. 160. Ikemura, T. (1985). Codon usage and tRNA content in unicellular and multicelular organisms. Molecular Biology and Evolution, 2, 13-34. Koubínová, D., Sreepada, K. S., Koubek, P., & Zima, J. (2010). Karyotypic Variation in Rhinolophid and Hipposiderid Bats (Chiroptera: Rhinolophidae, Hipposideridae). Acta Chiropterologica, 12(2), 393-400. Li, G., Jones, G., Rossiter, S.J., Chen, S.F., Parsons, S. & Zhang, S. (2006). Phylogenetics of small horseshoe bats from East Asia based on mitochondrial DNA sequence variation. Journal of Mammalogy, 87, 1234-1240. Maree, S., & Grant, W.S. (1997). Origins of horseshoe bats (Rhinolophus, Rhinolophidae) in southern Africa: evidence from allozyme variability. Journal of Mammalian Evolution, 4, 195–214. Maruyama, T., Gojobori, T., Aota, S.I., & Ikemura, T. (1986). Codon usage tabulated from the GenBank genetic sequence data. Nucleic Acids Research, 14, l51-197. McKenna, M.C., & Bell, S.K. (1997). Classification of Mammals above the Species Level. Columbia University Press. p. 631. Osawa, S., Ohama, T., Jukes, T.H., & Watanabe, K. (1989). Evolution of the mitochondrial genetic code. I. Origin of AGR serine and stop codons in metazoan mitochondria. 29(3), 202-207. Sengupta, S., Yang, X., & Higgs, P. G. (2007). The Mechanisms of Codon Reassignments in Mitochondrial Genetic Codes. Journal of Molecular Evolution, 64, 662-688. Simmons, N.B., & Geisler, J.H. (1998). Phylogenetic relationships of Icaronycteris, Archaeonyteris, Hassianycteris and Palaeochiropteryx to extant bat lineages, with comments on the evolution of echolocation and foraging strategies in Microchiroptera. The Bulletin of the American Museum of Natural History, 235, 1–182. Simmons, N.B. (2005). Order Chiroptera. In: Wilson, D.E., & Reeder, D.M. (Eds.), Mammal Species of the World: A Taxonomic and Geographic Reference, third ed. Smithsonian Institution Press, pp. 312–529. Stoffberg, S., Jacobs, D.S., Mackie, I.J., & Matthee, C.A. (2010). Molecular phylogenetics and historical biogeography of Rhinolophus bats. Molecular Phylogenetics and Evolution, 54, 1–9. 44 | P a g e �PROCEEDINGS th ______ The 5 International Symposium on Sustainable Development_______ ISSD 2014 Supek, F., & Vlahoviček, K. (2005). Comparison of codon usage measures and their applicability in prediction of microbial gene expressivity. BMC Bioinformatics, 6, 182. Supek, F., & Vlahoviček, K. (2006). INCA: synonymous codon usage analysis and clustering by means of selforganizing maps. BMC Bioinformatics, 20(14), 2329-2330. Teeling, E.C., Madeson, O., Van Den Bussche, R.A., De Jong, W.W., Stanhope, M.J., & Springer, M.S., (2002). Microbat paraphyly and the convergent evolution of a key innovation in Old World rhinolophoid microbats. PNAS USA, 99, 1431–1436. Teeling, E.C., Madsen, O., Murphy, W.J., Springer, M.S., & O’Brien, S.J. (2003). Nuclear gene sequences confirm ancient link between New Zealand’s short-tailed bat and South American noctilionoid bats. Molecular Phylogenetic and Evolution, 28, 308–319. Vaughan, T., Ryan, J., & Czaplewski, N. (2000). Mammalogy, 4th Edition. Toronto: Brooks Cole Press. Wilson, D., & Reeder, D. (2005). Mammal Species of the World, 3rd edition. Baltimore: Johns Hopkins University Press. Semir Dorić is postgraduate student at Faculty of Science, University of Sarajevo (Department of biology) with a special interest in field of molecular biology and bioinformatics. In diploma thesis has analyzed a codon usage in selected representatives of family Rhinolophidae (Chiroptera) with a dedication to implement different software packages towards improving the quality of his investigations. Lada Lukić Bilela is associated professor of Molecular biology and Genomics at Faculty of Science, University of Sarajevo (Department of biology). She received her BSc in Biology (Faculty of Science, Sarajevo) MS and PhD in Mitochondrial genomics of Porifera at (Faculty of Science, Zagreb; field: molecular and cell biology). She was employed in Ruđer Bošković Institute at Laboratory of molecular genetics and performed her postdoctoral training at Johannes Gutenberg University of Mainz (Marie Curries Training Network fellowship). 45 | P a g e � Dublin Core The Dublin Core metadata element set is common to all Omeka records, including items, files, and collections. For more information see, http://dublincore.org/documents/dces/. Extent The size or duration of the resource. 2451 Title A name given to the resource COMPARISON OF CODON USAGE IN MITOCHONDRIAL GENOMES OF RHINOLOPHID AND HIPPOSIDERID BATS Author Author DORIĆ, Semir LUKIĆ BILELA, Lada Abstract A summary of the resource. According to current phylogenetic hypotheses, the bats of the families Rhinolophidae and Hipposideridae are sister groups nested within the clade of Pteropodiformes. The Hipposideridae are family of bats commonly known as the Old World leaf-nose bats. While this family has long been considered as a rhinolophid subfamily Hipposiderinae, it is now more generally classified as its own family. The Hipposideridae contain 10 living genera and more than 70 species, mostly in the widespread genus Hipposideros. This study is an attempt to confirm a distinction between these two families by a codon usage comparison of a complete set of mitochondrial protein-coding genes from currently available mitochondrial (mt) genomes of rhinolophid and hiposiderid bats. The INCA 2.1 and GCUA 2.0 were used for the codon usage computing. Measure Independent of Length and Composition (MILC), was used to estimate the codon usage of 13 mt protein-coding genes from five species of genus Rhinolophus and one species of Hipposideros (while only four genes were available from H. larvatus). Large randomly generated sequence sets were used to test for dependence on (i) sequence length, (ii) overall amount of codon bias and (iii) codon bias discrepancy in the sequences. Our findings suggest no significant differences in codon usage bias, among analyzed rhinolophid species, by statistical estimation of absolute frequency values despite the changed MILC values for nd1 and nd3 from Hipposideros armiger. Keywords: MILC, MELP, bats, codon usage, codon frequencies Publisher An entity responsible for making the resource available International Burch University Date A point or period of time associated with an event in the lifecycle of the resource 2014-05-15 Keywords Keywords. Article PeerReviewed Identifier An unambiguous reference to the resource within a given context ISSN 978-9958-834-36-3 Q Science (General),QH301 Biology,QH426 Genetics https://omeka.ibu.edu.ba/files/original/966fa15171b70b16a0139ddd3f08f798.pdf 5bc17b869bc4211fc4b4cccd882de8f0 PDF Text Text Comparison of Decision Tree Methods for Intrusion Detection Fatih Ozturk, fozturk@ibu.edu.ba Abdulhamit Subasi International Burch University, Faculty of Engineering and Information Technologies, 71000, Sarajevo, Bosnia and Herzegovina. asubasi@ibu.edu.ba Abstract: The popularity of using Internet contains some risks of network attacks, and attack methods differ each day, thus information security problem has become a significant issue all over the world. Intrusion detection is one major research problem in network security, whose aim is to identify unusual access or attacks to secure internal networks. At the moment, it is an urgent need to detect, identify and prevent such attacks effectively. In this work, we compared efficiency of decision tree methods in intrusion detection system. We compared the accuracy, detection rate, false alarm rate for different attack types. Keywords: Decision Tree; CART (Classification and Regression Trees); ID3; C4.5; Random Forest; Internet attack; Intrusion detection system (IDS). 1. Introduction In today’s business oriented world information is very important. It is even considered as an intangible asset. The fastest way to bring the necessary information to end users is via Internet. Internet has been so deeply involved in our lives in a way that we have been dependent on it. It has even been used as an important component of business models (Shon & Moon, 2007). The internet has been used to bring the customers and the businesses together by applications such as websites and emails. This brings up a very important concern, the information security. Intrusion detection is one of the major research topics to prevent attacks from the internet for both companies and the end users. Firewalls may protect the networks but day by day the attacks became more complicated. Intrusion detection systems (IDSs) overcome this complexities providing ways to resist different types of suspicious network communications and computing habits. All of this is done assuming that the behaviour of intruders differs from an authenticated user (Stallings, 2006) (Tsai, et al. 2009). Generally IDSs divide into two main categories based on their detection methods: anomaly and misuse (signature) detection (Anderson, 1995) (Rhodes, Mahaffey, & Cannady, 2000). Deviation from normal operation can be flagged as intrusion with anomaly detection. Meanwhile the use of well-known attacks to the system can be caught by misuse detection (Tsai, et al. 2009). Many issues should be considered when building an IDS, like data collection, data pre-processing, intrusion recognition, reporting and response. The most important of these components is intrusion recognition (Wu, Banzhaf, 2009). In literature, since Denning first proposed the intrusion detection model in 1987 (Denning, 1987), different machine learning techniques are used to develop anomaly and misuse detection systems. Especially classifiers are used to detect whether a connection over internet is normal use or an attack (Wu, Banzhaf, 2009). Between late 80s and early 90s combination of expert systems and statistical methods were popular. Detection models have been acquired from field experts. From mid to late 90s normal abnormal detection moved to automatic modelling. Artificial intelligence (AI) and machine learning helped automation via a test data. Rule based induction, classification and data clustering highly referenced. Since IDS have huge network traffic volumes, highly diverse data distribution and difficult decision boundary to construct a good model is very challenging (Wu, Banzhaf, 2009). In this work, we compared different decision tree methods for intrusion detection using KDD’99 dataset. 401 �This paper is organized as follows. Section 2 provides some literature review on IDS systems. Section 3 describes the data set and performance evaluation used in this paper. Section 4 gives brief theoretical background for different decision tree algorithms. Section 5 gives our analysis results and we conclude in Section 6. 2. Literature review on intrusion detection system The IDS was first mentioned in Anderson’s technical report (Anderson, 1980) where he mentioned the use of statistical methods to analyse users’ behaviour to detect the misuse of the system. In 1987 Dorothy laid the foundations of Intrusion detection models. He proposed a prototype IDS: IDES (Intrusion Detection Expert Systems) (Denning, 1987). After this a number of IDS has been released such as Discovery, Haysack, MIDAS, NADIR, NSM, Wisdom and sense, DIDS, etc (Bace, 2002). (Wu, Yen, 2009). We can classify IDSs into two major areas: Misuse Detection and Anomaly detection (Bace, 2002). In misuse detection we try to match incoming data with pre-defined intrusive behaviour (signature). So, the well known intrusions are detected very fast and accurately (low false alarm rates). For this reason this method is adopted by many commercial IDSs. However the intrusion methods are not trivial and will evolve continuously. If an unknown attack comes to the system it will fail. To overcome this shortage we have to update the signature database at the expense of our valuable time (Wu, Yen, 2009). The second area, anomaly detection (Denning, 1987), may overcome this problem. The anomaly detection relies on the fact that most of the network communications are normal data and it tries to model this normal behaviour. Anything off this behaviour is marked as anomaly and a flag is raised by the system. Since the intrusions are rare and different from the normal data this method catches most of the abnormal behaviour even with unknown attacks. The difficulty here is the false alarm rates. The boundary between the normal and abnormal data is often too close (Wu, Banzhaf, 2009). Another problem faced is the constant changing of the normal usage statistics. For example we see more and more UDP connections since people use video sharing sites more frequently. There is also a hybrid method introduced by MINDS (Ertoz et al., 2004), EMERALD, Prelude, etc. where they try to leverage the disadvantages of both of these methods (Wu, Yen, 2009). 3. Datasets and performance evaluation The data we have used in the testing has been derived from well known KDD’99 datasets. This data is collected in 1998 by MIT Lincoln laboratory which has seven weeks of training and two weeks of test data. The test UNIX and Windows NT hosts and a Cisco router have faced more than 300 instances of 38 attack types. KDD’99 dataset is derived from this dataset in 1999 by assembling individual TCP packets into TCP connections. This data was the benchmark dataset used in the International Knowledge Discovery and Data Mining Tools competition (Tsai, et al. 2009). Each TCP connection has 41 features with a label specifying the status of the connection. The features consist of 38 numeric and 3 symbolic features falling into one of the following categories: 1- Basic features: 9 basic features to describe each individual TCP connection. 2- Content features: 13 domain knowledge related feature to indicate suspicious behaviour without any sequential patterns. 3- Time-based traffic features: 9 features used to summarize the connections in the past 2 s that had the same destination host or the same service as the current connection. 4- Host-based traffic features: 10 features were constructed using a window of 100 connections to the same host instead of a time window to see the attacks that might take more than 2 s. (Tsai, et al. 2009). In KDD’99 dataset we have 4,940,000 data instances covering the normal and 24 network attacks. The test set has 311,029 data instances with a total of 38 attacks, 14 of which do not appear in the training set. Because of this large training set, usually another dataset with the 10% of the data is used (Wu, Banzhaf, 2009). The attacks contained in test information can be separated into the following categories: Probe: These are not considered as real attacks but often used as preparation steps for a full scale attack. They are used to investigate the end system for future attacks. 402 �Dos (Denial of service): This type of attack usually keeps server busy or uses the valuable bandwidth so that it will not provide the necessary service to the end users. Most common types are SYN Flooding, Ping Flooding and etc. U2R (User to Root): In this type the attacker tries to take control of the admin user of the system from the leaks in the system. Buffer overflow is one of these methods. R2L (Remote to Local): It is used to take advantage of the server providing the services in order to get sensitive security information on the server or user personal files. Unicode leak, SQL injection and etc. are examples of this attack type (Yen, Wu, 2009). In our experiments we have selected a subset of 7000 instances to train our data. This subset has 6 different kinds of attacks and 1 normal data. The selected attack types are considered the most widely used types which include "port-sweep", "back-door", "IP-sweep", "nmap atack", "satan" and "smurf". The best method to evaluate the effectiveness of an IDS is the correct prediction ability. There are four possible outcomes according to the real nature of a sample data compared to the outcome (prediction) from the IDS. This is also known as the truth matrix. True negative (TN): The amount of normal data predicted when it is really normal, True positive rate (TP): The amount of attack predicted when it is really attack False negative rate (FN): The amount of normal predicted when it is really attack False positive rate (FP): The amount of attack predicted when it is really normal. TN/(TN+FP) gives specificity or False alarm rate (FAR), TP/(TP+FN) gives sensitivity or detection rate (DR), The most widely used performance evaluation are Detection Rate with False Alarm Rate. A good IDS must have high DR and a low or zero FAR (Wu, Yen, 2009) (Wu, Banzhaf, 2009). 4. Decision trees In this paper we have compared four different Decision Tree methods. In a decision tree we try to classify a sample through some decisions leading us to a succeeding decision. The classification takes places from the root node to the leaf where the end leaf has the category information. Each node holds one attribute of the tree and the branches corresponds to the value of that attribute. (Mitchell, 1997). It is pretty much similar to flow chart structure; the test properties are internal nodes, test results corresponds to each branches, and distribution situation of various types are nodes of leaves. Each decision tree will belong to one of these categories: top-down tree construction or bottom-up pruning. The most essential and common method used for classification tree is CART (Classification and Regression Trees) (Breiman, Friedman, Olshen, & Stone, 1984), ID3 and C4.5 (Quinlan, 1993) and they are all top-down tree construction. The common algorithm can be described as such; (i) place all training set into the root of classification, (ii) check whether it contains all the same type or an empty set; if node contains more than one type of training set check each property of data according to certain function, and select a proper property. Divide training set into N parts, each constituting a new node to the root node according to the value of the property. This process is called as splitting node. (iii) Check whether each node is a leaf; if not, split them into new nodes as described in "ii". (iv) Proceed with splitting until each nodes turn into a leaf. This will construct our tree. In these methods we can totally classify any given training data into branches and leaves of the tree (Yen, Wu, 2009). In ID3 method, by definition, all the parameters given should be discrete values. In our sample set we have 3 symbolic features. In order to overcome this problem we have converted these values into discrete valued properties with a pre-process before evaluating our sample data. Under certain conditions Decision Trees have advantage over other common supervised learning methods like discriminant analysis. Especially they do not suffer the same probability distribution restrictions. There is also no necessity to assume linear model and they are very useful with non-linear predictors (Wu, Banzhaf, 2009). 4.1 CART In CART (Classification and Regression Trees) there are six general questions to be answered: 1- Should we allow properties to be restricted to binary values or let the multi valued properties exist? 2- Which property should be tested at a node? 3- When do we declare a node as a leaf 403 �4- How to make trees smaller and simpler when it becomes too large? Prune? 5- When we induce an impure leaf node how shall we label it? 6- How can we handle the missing data? According to these questions we create a tree giving good enough results with easy to compute and fast responses (Duda, Hart & Stork, 2002). 4.2 ID3 Since this is the third in a series of identification or "ID" process it is called ID3. The inputs were intended to be used with nominal (unordered) inputs only. When a real-valued variable is present it is first divided into intervals and then each interval applied as nominal input. Here each split has a branching factor Bj, where it is the number of discrete attribute container of variable j chosen for splitting. Usually these are not binary and a gain ratio impurity should be used. The number of levels they have is equal to the number input variables. It continues to run until all nodes are pure and no more possible splits exist. If necessary a common pruning technique can be applied to the algorithm (Duda, Hart & Stork, 2002). 4.3 C4.5 This is the successor and refinement of ID3 method, and very popular among classification tree methods. The real valued variables are handled as in CART. Nominal data is used to crate multi-way splits as in ID3 with a gain ratio impurity. The pruning is achieved with the statistical significance of the splits. The main difference between CART and C4.5 is the missing features. There are no substitute splits precomputed. If a defective test pattern with missing feature is present at branch N with branching factor B, C4.5 follows all possible B answers to the descendent nodes and at last to the B leaf nodes. The final decision is made according to the labels of B leaf nodes multiplied by the decision probability of N. Here, unlike CART, we don't exploit statistical correlations between different features of the training points. There is no extra computation and storage required for C4.5. So, it is much preferred where the storage is a major concern. As for the pruning, C4.5 generates the rules from the tree where each leaf node has one associated rule (the route from the root node to that leaf). Then it deletes the redundant antecedents in these rules (Duda, Hart & Stork, 2002). 4.4 Random forests Decision Trees can be less accurate than methods like support vector machines and the structure of the trees can also be unstable (Breiman, 1996). Small changes in the training data may significantly result in changes to the tree, either to the identity of the split variable or to the value of the split. Even the structure changes significantly the resulting prediction may stay the same. RandomForest (a trademark of Saldorf Systems) algrithm is developed by Breiman (Breiman, 2001a,b ) to overcome these shortcomings while possibly enhancing the interpretability. Some of the important features of RandomForest is: (Breiman and Cutler, 2004b) - The accuracy that equals or tops many current classifiers without overfitting - Fast on large databases and handle thousands of predictors without selector routines - Each predictor importance is estimated - Generalization error estimate is generated in an unbiased manner - The missing data and error balance when a class proportion marked differently is dealt with robust algorithms. - The proximities between the pairs of cases that can be used in clustering and identifying outliers are computed - Variable interaction detection is done by an experimental method - Generated forest can be saved to be used on other data. The forest, generated by many trees, where each tree is different from each other at the generation of training cases and predictors used at each node. Each tree is generated by a subset (m) of available predictors drawn randomly, which is much less than the total available. This m value is the only adjustable parameter which the random forests are sensitive (Breiman and Cutler, 2004a). This parameter is the same for all the trees in forest. Each tree is allowed to grow to the most possible extend thus; there is no need for pruning. Besides increased prediction accuracy, they help to determine the importance of each variable and associations between each case (Fielding, 2007). 404 �5. Results and Discussion Over the past decade intrusion detection based on machine learning methods has been extensively studied topic, and they satisfy the growing demand of reliable and intelligent intrusion detection systems. In this study we compared the performance of different decision tree classifiers on solving intrusion detection problems. This research works were trained and tested on the KDD’99 dataset. The classification results are shown in Table 1. From this table, we can easily see that there is no significant difference between accuracy of these methods; however, random forest is better than others. According to the average, ID3 is worst in the classification. Accuracy refers to the proportion of attack detected among all attack data, namely, the situation of TP. In detection rate, the random forest has accuracy 99.94 % approximately. False alarm rate refers to the proportion that normal data is falsely detected as attack behaviour, namely, the situation of FP. In comparison of false alarm rate, random forest is 0.1 %, but it is worse than C4.5 and CART. According to the average value, false alarm rate of C4.5 and CART is 0 % and better than random forest. Decision Tree Method ID3 CART C4.5 Random Forest Accuracy 97.4% 99.8286 % 99.8857 % 99.9429 % Table 1: Classification of Different Decision Tree Algorithms We have considered the problem of comparing different decision tree classifiers, including ID3, CART, C4.5 and random forest. Here, rather than directly comparing typical implementations of CART, ID3, C4.5 and random forest methods, it is more useful to consider distinctions within the different component steps. Anybody can build a tree using any practical feature processing, impurity measure, stopping criterion or pruning method. Of course, if the designer has insight into feature pre-processing, this must be utilized. Generally, pruning can be preferred over stopped training and cross-validation, because it takes advantage of more of the information in the training set. On the other hand, pruning large training sets can be computationally expensive. The pruning of rules is less useful for problems that have high noise and are at base statistical in nature. Similar to the most classification methods, one gains expertise and insight through experimentation on a wide range of problems. Any single tree algorithm neither dominates nor is dominated by other classification methods. It can be seen that trees yield classifiers with accuracy as good as other classification methods. (Duda, Hart & Stork, 2002) Even though some promising results have been accomplished by different decision tree classifier to IDSs, there are still challenges that lie ahead for researchers in this area. First and foremost, good benchmark datasets for network intrusion detection are needed. The KDD’99 is the most important benchmarks used to evaluate the performance of network intrusion detection systems. But, they are suffering from a serious drawback: failing to realistically simulate a real-world network (Brugger 2007)(Mahoney, Chan, 2003) (McHugh, 2000). An IDS working well on these datasets may demonstrate unacceptable performance in real environments. (Wu, Banzhaf, 2009) These datasets possess some special characteristics, such as huge volume, high dimension and highly skewed data distribution. As a result, using only these datasets is not adequate to demonstrate the efficiency of a learning algorithm. It is also meaningful to note that the KDD’99 datasets were collected about 10 years ago. One of the important characteristics of intrusion detection is the capability of adaptation to continually changing environments. Not only the intrusive behaviour evolves continuously, but also the legitimate behaviour of users, systems or networks changes over time. If the IDS is not flexible enough to cope with behavioural changes, detection accuracy will dramatically decrease. A focus on adaptation in IDSs is highly recommended. Another challenge to confront in IDS is the huge volume of audit data that makes it difficult to build an effective IDS. Perhaps it is time to create a new and high-quality dataset for the intrusion detection task. (Wu, Banzhaf, 2009) 6. Conclusion Intrusion detection based on computational intelligence is currently attracting considerable interest from the research community. This research compares accuracy, detection rate and false alarm rate of different attacks. KDD’99 dataset is current benchmark dataset in intrusion detection. For comparison results of decision tree 405 �algorithms, we find that Random forest is superior to others in accuracy and detection; ID3 is the worst. In comparison of false alarm rate, C4.5 and CART are better than Random forest. Through test and comparison, the accuracy and detection rate of Random forest is higher than that of others, but false alarm rate of C4.5 and CART is better; if we combine the two methods, overall accuracy can be increased greatly. Dataset KDD’99 applied in the research is popularly used in current intrusion detection system; however, it is data of 1999, and network technology and attack methods changes greatly, it cannot reflect real network situation nowadays. Therefore, if newer information is got and tested and compared refresh, they can more accurately reflect current network situation. References: Anderson, James P. (1980). Computer security threat monitoring and surveillance, technical report, James P. Anderson Co., Fort Washington, Pennsylvania. Anderson, J. (1995). An introduction to neural networks. Cambridge: MIT Press. Bace, Rebecca G. (2002). NIST special publication on intrusion detection systems. Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, P. J. (1984). Classification and regressing trees. California: Wadsworth International Group. Breiman, L. (1996). Bagging predictors. Machine Learning, 24, 123_40. Breiman, L. (2001a). Random forests. Machine Learning, 45, 5_32. Breiman, L. (2001b). Statistical modelling: the two cultures. Statistical Science, 16, 199_215. Breiman, L. and Cutler, A. (2004a). Interface Workshop: April 2004. Available at http://statwww.berkeley.edu/users/breiman/RandomForests/interface04.pdf (accessed 1 February 2006). Breiman, L. and Cutler, A. (2004b). Random Forests. Available at http://statwww.berkeley.edu/users/breiman/RandomForests/cc_home.htm (accessed 1 February 2006). Brugger, T. (2007) KDD cup’99 dataset (network intrusion) considered harmful, 15 September 2007. Retrieved January 26, 2008, from http://www.kdnuggets.com/news/2007/n18/4i.html. Dorothy, Denning. 1987. An intrusion detection model. IEEE Transaction on Software Engineering. Intrusion detection by machine learning: A review Duda R. O.; Hart, P. E. and Stork D. (2002). Pattern Classification, 2nd. Edition, John Wiley & Sons, 2002. Rhodes, B., Mahaffey, J., & Cannady, J. (2000). Multiple self-organizing maps for intrusion detection. In Paper presented at the proceedings of the 23rd national information systems security conference. Baltimore, MD. Ertoz, L., Eilertson, E., Lazarevic, A., Tan, P., Srivastava, J., Kumar, V., et al. (2004). The MINDS – minnesota intrusion detection system. Next generation data mining. MIT Press. Fielding, Alan H. (2007). Cluster and Classification Techniques for the Biosciences, Cambridge University Press The Edinburgh Building, Cambridge cb2 2ru, UK, 2007. Mahoney, M.V.; Chan, P.K. (2003). An analysis of the 1999 DARPA/Lincoln laboratory evaluation data for network anomaly detection. Technical Report TR CS-2003-02, Computer Science Department, Florida Institute of Technology, 2003 McHugh, J. (2000). Testing intrusion detection systems: a critique of the 1998 and 1999 darpa intrusion detection system evaluations as performed by Lincoln laboratory, ACM Transactions on Information and System Security 3 (4) (2000) 262– 294. Mitchell, T. (1997). Machine learning. New york: McGraw Hill. Quinlan, J. R. (1993). C4.5: Programs for machine learning. Morgan Kaufmann Publishers. Shon, T., & Moon, J. (2007). A hybrid machine learning approach to network anomaly detection. Information Sciences, 177, 3799–3821. 406 �Stallings, W. (2006). Cryptography and network security principles and practices. USA: Prentice Hall. Tsai, Chih-Fong; Hsu, Yu-Feng; Lin, Chia-Ying; Lin, Wei-Yang (2009). Intrusion detection by machine learning: A review. Expert Systems with Applications, 36 (2009) 11994–12000 Wu, Shelly Xiaonan; Banzhaf, Wolfgang (2009). The use of computational intelligence in intrusion detection systems: A review. Applied Soft Computing, 10 (2010) 1–35 Wu, Su-Yun; Yen, Ester (2009). Data mining-based intrusion detectors. Expert Systems with Applications, 36 (2009) 5605– 5612 407 � Dublin Core The Dublin Core metadata element set is common to all Omeka records, including items, files, and collections. For more information see, http://dublincore.org/documents/dces/. Extent The size or duration of the resource. 535 Title A name given to the resource Comparison of Decision Tree Methods for Intrusion Detection Author Author Ozturk, Fatih Subasi, Abdulhamit Abstract A summary of the resource. The popularity of using Internet contains some risks of network attacks, and attack methods differ each day, thus information security problem has become a significant issue all over the world. Intrusion detection is one major research problem in network security, whose aim is to identify unusual access or attacks to secure internal networks. At the moment, it is an urgent need to detect, identify and prevent such attacks effectively. In this work, we compared efficiency of decision tree methods in intrusion detection system. We compared the accuracy, detection rate, false alarm rate for different attack types. Date A point or period of time associated with an event in the lifecycle of the resource 2010-06 Keywords Keywords. Conference or Workshop Item PeerReviewed Q Science (General) https://omeka.ibu.edu.ba/files/original/97f80968ecbf84e9fa61397f3c428f07.docx cae459a3db56f9616e8ce14714225e07 https://omeka.ibu.edu.ba/files/original/01e92947115f5804971b587bd676a6c2.pdf dc6ccd9131157e9f3d547dbbbde8e07c PDF Text Text COMPARISON OF ELECTROMAGNETIC RADIATION LIMITS FOR EXTREMELY LOW FREQUENCIES IN EUROPEAN COUNTRIES Mahmut Yalçın Istanbul University, Istanbul, Turkey myalcin@Istanbul.edu.tr Keywords:Electromagnetic Radiation (EMR); Electromagnetic Pollution; Extremely Low Frequency (ELF); Exposure Limits; Magnetic Field; High Power Lines. ABSTRACT Almost every member of modern societies constantly live in electromagnetic fields (EMF) which are much higher than those found in nature. Power lines, computer monitors, different electrical equipments, radio, television, mobile phones, microwave ovens can be given examples of these EMF sources. Their potential effects of health continue to be the subject of controversy. Extremely low frequency (ELF, 0-3000 Hz) region of spectrum is radiated by transformers, household equipments, high power lines, and by electrical goods is investigated in this study. Since more than 25 years research efforts to find a correlation between the electromagnetic field and their effects on health of human are going on, but without significant success. Generally, countries accept the standarts of International Commission on Non-Ionizing Radiation Protection (ICNIRP), World Health Organization (WHO), and European Committee for Electrotechnical Standardization (CENELEC). Some countries apply more strict limit values than above foundations. The International Agency for Research on Cancer (IARC) reviewed EMFs and cancer in June 2001, and classified magnetic fields as “possibly” carcinogenic for low frequency region. Therefore, exposure limit values of EMFs are really important. The best way is acceptance of As Low As Reasonably Achievable (ALARA) principle as long as have not exact scientific results. � Dublin Core The Dublin Core metadata element set is common to all Omeka records, including items, files, and collections. For more information see, http://dublincore.org/documents/dces/. Extent The size or duration of the resource. 2131 Title A name given to the resource COMPARISON OF ELECTROMAGNETIC RADIATION LIMITS FOR EXTREMELY LOW FREQUENCIES IN EUROPEAN COUNTRIES Author Author YALCIN, Mahmut Abstract A summary of the resource. Keywords:Electromagnetic Radiation (EMR); Electromagnetic Pollution; Extremely Low Frequency (ELF); Exposure Limits; Magnetic Field; High Power Lines. ABSTRACT Almost every member of modern societies constantly live in electromagnetic fields (EMF) which are much higher than those found in nature. Power lines, computer monitors, different electrical equipments, radio, television, mobile phones, microwave ovens can be given examples of these EMF sources. Their potential effects of health continue to be the subject of controversy. Extremely low frequency (ELF, 0-3000 Hz) region of spectrum is radiated by transformers, household equipments, high power lines, and by electrical goods is investigated in this study. Since more than 25 years research efforts to find a correlation between the electromagnetic field and their effects on health of human are going on, but without significant success. Generally, countries accept the standarts of International Commission on Non-Ionizing Radiation Protection (ICNIRP), World Health Organization (WHO), and European Committee for Electrotechnical Standardization (CENELEC). Some countries apply more strict limit values than above foundations. The International Agency for Research on Cancer (IARC) reviewed EMFs and cancer in June 2001, and classified magnetic fields as “possibly” carcinogenic for low frequency region. Therefore, exposure limit values of EMFs are really important. The best way is acceptance of As Low As Reasonably Achievable (ALARA) principle as long as have not exact scientific results. Publisher An entity responsible for making the resource available International Burch University Date A point or period of time associated with an event in the lifecycle of the resource 2013-05-24 Keywords Keywords. Article PeerReviewed Identifier An unambiguous reference to the resource within a given context ISSN 2233-1565 https://omeka.ibu.edu.ba/files/original/88f415b31cdd1dd53a6ff7a143eaf0f9.pdf d00733f9a46e03b44d78ea56bf2948a9 PDF Text Text Comparison of Eleven Mathematical Models for describing the first Lactation Curve of Holstein Cattle in Turkey Đsmail Keskin Department of Animal Science, Faculty of Agriculture, Selcuk University, 42075, Konya, Turkey ikeskin@selcuk.edu.tr Nazire Memmedova Department of Animal Science, Faculty of Agriculture, Selcuk University, 42075, Konya, Turkey naziramamedova@yahoo.com Fatma Đlhan Department of Animal Science, Faculty of Agriculture, Selcuk University, 42075, Konya, Turkey fatmailhan@selcuk.edu.tr Birol Dağ Department of Animal Science, Faculty of Agriculture, Selcuk University, 42075, Konya, Turkey bdag@selcuk.edu.tr Fariz Mikailsoy Department of Soil Science and Plant Nutrition, Faculty of Agriculture, Selcuk University, 42075, Konya, Turkey farizm@selcuk.edu.tr Abstract: In this study, eleven standard lactation curve models (Incomplete Gamma (WD), Quadratic (Q), Cubic (C), Linear Hyperbolic Function (LH), Inverse Polynomial Function (IP), Mixed Log (MIL), Exponential (WIL), Dhanoa (DH), Cobby and Le Du (CD), Polynomial Regression (AS) and New Model (NM)) were used to predict a typical dairy cow lactation derived as the average daily milk yield of 105 complete first lactations of HolsteinFriesian cows in one herd. Milk yield controls were made daily in this farm. Total milk yield (TMY) was calculated from observation of daily milk yield. TMY was also predicted by using 11 different models. The total milk yields predicted by the models were very close to each other and the differences between observation of TMY were not found statistically significant (P>0.05). The models were found to be adequate for estimation of milk yield. Determination coefficients (R2) of the models ranged from 67.15 % to 86.68 %. In comparison of the models, the TMY, peak yield (PY), peak time (PT), persistency (P), mean square prediction error (MSPE), approximation error ( ε ), reliability criterion for estimating trustworthiness of the determination coefficient ( σ ), standard error ( θ ) and Durbin-Watson (DW) values were evaluated together. Using the AS, WD and new developed NM models accurately predicted the milk yield of Holstein cows. Key Words: Holstein, Cows, Lactation Curve, Milk Yield, Mathematical Model 246 �Introduction Turkey has 11.3 million head of cattle and 70 % of them are improved cattle and their crossbreeds. The number of milked animals is 4.2 million and approximately 3 million of them are improved cattle and their crossbreeds. The milk obtained from cows is 11.3 million tons and nearly 86 % of the milk production is produced by improved cattle and their crossbreds. But the lactation milk yields is very low (i.e. for native breeds 1.3, for crossbreeds 2.7 and for improved cattle 3.9 tons) (TurkStat, 2007). Producers aim to increase milk yield and decrease cost for a profitable dairy cattle production. Persistency is one of the most important factors which determine milk production cost along the lactation. Milk yield begins with calving and reach to highest level between 40 to 70 days and then continues to decrease along the lactation. With decreasing of daily milk yield, the production cost begins to increase from day to day (Gengler, 1996; Koçak and Ekiz, 2006). A mathematical model of the lactation curve provides summary information about dairy cattle production, which is useful in making management and breeding decisions and in simulating a dairy enterprise (Olori et al., 1999). In order to asses plausible forms of lactation curves, milk yield records collected throughout the whole lactation are required. But most of the small and medium sized dairy farms in Turkey still use classical milking systems. Milk yield is generally recorded monthly in these farms. The lactation curve models enable them to evaluate lactation as a whole. So that the lactation curve shape is determined and unbiased comparison methods among animals with incomplete lactation records for genetic evaluation purposes can be practiced (Keown and van Vleck, 1973). Knowledge of the lactation curve allows prediction of total milk production from partial production measured at several test days early in lactation (Goodal and Sprevak, 1985). Animals with a high milk yield potential can be identified by using this information before the whole lactation is completed. Also, lactation curves can be used for prediction of lifetime milk production from early lactation traits (Dalal et al., 2004), culling, assessing nutritional and health status of animals (Duoduet, 1982; Souvant and Fehr, 1975) and evaluating a suitable time to end milking (Chang et al., 2001). The first mathematical model aimed to describe lactation curves was developed by Brody et al., (1923). It was notified that this model was followed by the models reported out by Sika (1950), Nelder, (1966), Wood (1967), Dave (1971) and Jenkins and Ferrel (1984) (Landete-Castillejos and Gallego, 2000). The Wood model has been used in most lactation curve model studies, because it includes the basic features of lactation curves with only three parameters a, b and c which allow the calculation of average yield, peak yield and peak time, respectively. This has made the Wood model the most widely used function for the description of lactation curves. Most of the alternative models are also based on the Wood model (Cobby and Le Du, 1978; Wilmink, 1987; Papajcsik and Bodero, 1988). However, some mathematical models have been proposed to describe the regular shape of the lactation curve in dairy cows from partial or incomplete data (Neal and Thornley, 1983; Goodal and Sprevak, 1984; Batra, 1986; Morant and Gnanasakthy, 1989; Dijkstra et al., 1997; Olori et al., 1999; Vargas et al., 2000). Also these models provide analysing systemic changes in milk yield caused by environmental factors (Goodall and Sprevak, 1985; Morant and Gnanasakthy, 1989) and determining the milk production characteristics such as persistency (Gengler, 1996), peak yield and time to peak yield (Masselin et al., 1987; Gipson and Grossman, 1990). The objective of this study was to compare the suitability of WD, WIL, MIL, C, Q, DH, IP, CD, LH, AS and NM models to the first lactation data of Holstein cows. Materials and Methods The data of this study was from the first lactation records of 105 Holstein cows raised in a private enterprise in the Karapınar district (37o 42' K, 33o 35' D and 994 m above sea level) of the Konya Province in the Central Anatolia Region of the Turkey. The data were collected from the first lactation records of the cows that gave birth in 2004. They were machine milked twice daily and milking records were started 3th days of lactation. There is a computer-based herd managing program in the enterprise and milk yield controls were made daily. Average lactation length was 312±4.37 days. The experiment was carried out according to guidelines of Selçuk University Faculty of Agriculture located in the Konya Province. In the study, to explain lactation curves, eleven different empirical mathematical models were used together and compared. These models are as follows: (1) Incomplete Gamma (WD), (Wood, 1967): Y( t ) = at b e − ct (2) Quadratic (Q), (Dave, 1971): Y( t ) = a + bt + ct 2 (3) Cubic (C), 247 �Y( t ) = a + bt + ct 2 + dt 3 (4) Exponential (WIL), (Wilmink, 1987): Y( t ) = a + be − kt + ct (Which was fitted with the parameter k fixed at 0.61) (5) Mixed log (MIL), (Guo and Swallve, 1995): Y( t ) = a + bt 1 / 2 + c log t (6) Polynomial Regression (AS), (Ali and Schaeffer, 1987): Y( t ) = a + bt + ct 2 + d log t + e log t 2 (7) Cobby and Le Du (CD), (Cobby and Le Du, 1978): Yt = a − bt − ae − ct (8) Linear Hyperbolic Function (LH), (Bianchini, 1984): Y( t ) = a + bt + c (1 / t ) (9) Inverse Polynomial Function (IP), (Nelder, 1966): Y( t ) = t /(a + bt + ct 2 ) (10) Dhanoa (DH), (Dhanoa, 1981): Yt = at bc e ( − ct ) (11) New developed model (NM): Y( t ) = at b e − ct − d t For all models, Yt is the observed milk yield at day t, a: is linked to milk yield at the beginning of lactation, b: to the ascending phase before peak yield, c: to the decreasing phase after peak yield, d: parameters which characterize the shape of the curve e: is the base of natural logarithm, which were estimated from a nonlinear regression analysis using the Statistica program. The WIL model has a total of four parameters, with k being exponent; following Wilmink (1987) a fixed value of k was used, which was estimated at 0.61 in a preliminary analysis as the best fitting value for the herd mean data. Subsequently the WIL model was considered as a three parameter curve in analysis of individual animals. Persistency (P) was calculated as: k P (%) = ∑ (pi + 1) / pi i =1 k × 100 Where pi is the yield of the record i that starts at peak time and k is the record number from peak time to the end of lactation (Sturtevant, 1986). The parameters obtained were used to calculate the predicted yields in the original equations at above. Residuals, defined as the absolute values of the difference between the predicted yield and real data of daily milk yield, were calculated and then the mean square prediction error (MSPE) for each lactation curve fitted was calculated and averaged for each model (Ruiz et al., 2000). True peak yield (PY) was determined from the test day milk yield means for the 105 cows and true peak time (PT) was determined as the average day on which daily milk yields were at their maximum value. Peak time values of the models were calculated by equalizing the first partial derivations of the functions to zero, and PY values were found for each cow by replacing PT values in the functions. The Durbin-Watson statistics was used as a measure of first order positive autocorrelation to test whether the residuals were randomly distributed (Grossman and Koops, 1988). DW was calculated for each lactation and models. Approximation error was calculated as: ε= 100 n Yi − Y ∑ n i =1 Yi Reliability criterion for estimating trustworthiness of the determination coefficient was calculated as: 248 �θ= R2 n 1 − (R 2 ) 2 Standard error was calculated as: n ∑ (Y i σ= − Y) i =1 n−m The models were compared in respect of their MSPE, correlation between yields and residuals (RESC), 2 R , TMY, peak yield (PY), peak time (PT), persistency (P), ε , θ and σ . Results and Discussion Lactation curve parameters in Holstein cattle were given in Table 1. The parameter a, which expressed the milk yield at beginning was 0.10 in IP model, and ranged from 12.71 to 25.81 among the other models. Estimates of parameter a, were found in this study more higher than estimated for WD, Q, C, LHF and IP models in Brown Swiss cows by Keskin and Tozluca (2004), for Q, C and WIL models in Simmental cows by Çilek and Keskin (2008), for WD, IP and AS models in Holstein-Friesian cows by Olori et al., (1999); but less than estimated for WD, MIL and AS models in Simmental cows by Çilek and Keskin (2008), for WD, MIL, WIL and AS models in Brown Swiss cows by Keskin et al., (2009), for MIL and WIL models by Olori et al., (1999) and very close to value estimated for WIL model in Holstein cows by Dědková and Němcová (2003). Models* Model parameters a ± Sa b ± Sb WD Q 15.34±0.428 22.04±0.419 0.161±0.0081 0.024±0.0038 C 20.53±0.649 AS 12.71±0.474 WIL 24.96±0.477 MIL 13.25±0.483 DH 20.38±0.759 LHF 25.81±0.438 0.062±0.0124 0.640±0.1099 26.064±1.1519 1.886±0.0831 49.468±11.7626 0.034±0.0015 0.032±0.0015 c ± Sc 0.0030±0.00012 0.0002±0.00001 -0.0004±0.00008 2.1894±0.34652 d ± Sd 0.00000054±0.000 0001 6.330±0.5275 e ± Se 1.022±0.1533 0.0307±0.00140 6.2888±0.25894 0.0020±0.00017 17.1840±0.74544 CD 25.44±0.439 18.3727±3.3172 0 IP 0.10±0.012 0.034±0.0009 0.0001±0.00000 5 NM 17.11±0.704 0.145±0.0118 0.0029±0.00013 0.153±0.0587 * : WD: Incomplete Gamma, Q: Quadratic, C: Cubic, LH: Linear Hyperbolic Function, IP: Inverse Polynomial Function, MIL: Mixed Log, WIL: Exponential, DH: Dhanoa, CD: Cobby and Le Du, AS: Polynomial Regression and NM: New Model Table 1. Estimates of the model parameters and their standard errors of eleven models The highest estimate of parameter b was fixed in WD model, but the lowest estimate was fixed in DH model. Estimates of parameter c were ranged from -17.1840 to 18.3727. Estimated lactation curve parameters for this herd were generally different from the previous studies. It may be due to raising in different environmental conditions or a result of management and administration in different ways in terms of milk production. On the other hand it is well known that the Holstein breed is more productive in temperate climatic zones and its milk production capacity may change by geographical regions. 249 �The parameter a expressing milk yield at the beginning and the b parameter indicating the speed of curve increase must be higher and the c parameter meaning the speed of curve decrease must be lower in order to obtain more milk production. The lactation curves of Holstein cows are given in Figures 1. As seen in this figure, fit lines of WD, AS and NM models are very close to the observed values. The total milk yield predicted by different models is very close to observed total milk yield, and the differences between them, were not significant (P>0.05). 250 �MSPE 5.95±0.321 6.11±0.300 5.75±0.712 4.90±0.256 7.04±0.370 5.99±0.329 7.90±0.542 6.57±0.356 6.88±0.370 7.57±0.418 5.67±0.315 RESC -0.13 0.30 0.31 0.00 0.17 -0.17 0.61 0.04 0.13 -0.35 -0.17 Ɛ 14.00 13.71 14.12 12.74 15.33 14.29 15.90 14.95 15.13 15.78 13.93 σ 2.47 2.43 2.55 2.18 2.59 2.39 2.69 2.51 2.57 2.70 2.33 Ɵ 45.52 39.61 41.70 54.36 34.74 44.71 36.90 39.01 36.77 31.69 47.30 DW 0.849 0.776 0.804 0.949 0.716 0.833 0.737 0.775 0.732 0.673 0.877 Table 2. Comparison of the models for estimating total milk yield (TMY), peak yield (PY), time to peak yield (PT), Persistency (P), Correlation between yields and residuals (RESC) and goodness-of-fit statistics (R2 and MSPE values) Goodness of fit statistics Models TMY (l) PY (l) PT (day) P (%) R2 (%) ns bc a ns WD 6407±150 24.50±0.368 72.84±17.20 99.7±0.01 76.17±0.014 Q 6370±148 ns 22.04±0.414d 0.00±0.000d 99.5±0.01 ns 75.23±0.013 20.53±0.649d 0.00±0.000d 99.6±0.03 ns 77.87±0.011 C 6338±145 ns 23.96±0.419c 56.14±3.267a 99.8±0.03 ns 80.65±0.011 AS 6370±148 ns 24.90±0.448bc 10.16±0.15c 99.9±0.05 ns 86.68±0.171 WIL 6370±148 ns 25.29±1.078bc 62.79±10.82a 99.8±0.04 ns 75.25±0.015 MIL 6370±148 ns 24.46±0.358bc -46.17±11.66d 99.6±0.01 ns 63.59±0.025 DH 6372±148 ns 24.71±0.366bc 23.69±0.771c 99.9±0.38 ns 72.22±0.016 LHF 6370±148 ns 25.08±0.395bc 13.40±1.288c 99.8±0.03 ns 69.93±0.019 CD 6364±148 ns IP 6335±147 ns 22.88±0.424b 29.31±1.287bc 99.9±0.02 ns 67.01±0.019 24.65±0.366bc 45.67±2.89ab 99.6±0.03 ns 76.88±0.014 NM 6309±159 ns 28.72±0.422a 69.38±4.85a 99.9±0.03 ns Really 6369±149 ns a, b, c, d : The means within columns with different superscript are significantly different at P<0.01 ns : not significant. �30 25 20 Milk Yield (l) 15 10 5 0 1 16 31 46 61 76 91 106 121 136 151 166 181 196 211 226 241 256 271 286 301 Day of Lactation Data AS WD WL IP LH MIL NM DH Q C CD Figure 1. Shape of lactation curve according to the models Total milk yield (TMY), peak yield (PY), time to peak yield (PT), Persistency (P), Correlation between yields and residuals (RESC) and goodness-of-fit statistics (R2 and MSPE values) of the models are given in Table 2. In this study, the differences between estimated and observed peak yields were found significant (P<0.01). But the differences between Q and C models’ peak yields were not significant and very close predictions obtained with WD, WIL, MIL, DH, LHF, CD, IP and NM models. However the really peak yield was found as 28.72, all models underestimated the peak yield. Generally the curves of real lactation data based on daily milk yields are very fluctuant. But the estimated yields of the models are not fluctuant and also the estimated peak yields do not show sharply increases as seen in real data. On the other hand it is expected that the real lactation data of well managed herds do not fluctuate. The peak time obtained from actual milk yield and the predicted PT from WD, AS and MIL models were very close and the differences between them, were not significant (P>0.05). But, the peak time of DH model was negative and the peak times of Q and C models were estimated to be only zero. It is likely caused by the decreasing curves from the beginning to the end of lactation for DH, Q and C models estimated by present data as seen in Figure 1. Persistency (P) values among all models were very close to each other and the differences between them, were not significant (P>0.05). The P values, were found in this study were higher than values predicted for WD, Q, C, LHF and IP models in Brown Swiss cows by Keskin and Tozluca (2004), for Q, C, WD, MIL, AS and WIL models in Simmental cows by Çilek and Keskin (2008). Higher determination coefficients for the used models show good fitting level of independent variables for explaining dependent variables. For all models R2 value were estimated between 63.59 % and 86.68 %, it was obtained the lowest value in DH (63.59) model, the highest in WIL (86.68) model. The best fitness was obtained with WIL, it was followed by AS, but DH fitted worst. The lowest MSPE values were fixed in AS model, then in WD model. The highest MSPE value was fixed in DH model. The R2 in this study, were lower than R2 values, were notified for WD, Q, C, LHF and IP models in Brown Swiss cows by Keskin and Tozluca (2004), for Q, C, WD, MIL, AS and WIL models in Simmental cows by Çilek and Keskin (2008), for WD, IP, WIL, MIL and AS models in Holstein-Friesian cows by Olori et al., (1999). MSPE values were higher than values, notified for Q, C, WD, MIL, AS and WIL models in Simmental cows by Çilek and Keskin (2008). 252 �The lowest ε values were realized in AS model and it was followed by Q, NM and WD models. The same condition was found for σ values, too. The highest θ values were estimated in AS model, it was followed by NM and WD models. Autocorrelation values for all models were close to zero, indicating positive autocorrelation which may pose problems with statistical inferences about the models. Correlations between the residuals and observed milk yield (RESC) ranged between -0.35 (IP) to 0.61 (DH) for all models. Though estimated residuals generally increased with observed yields, there are a little except for C, Q and DH. High daily yields being most difficult to predict while very low yields also caused problems (Olori et al., 1999). Conclusion The TMY, PY, PT and P values in AS, WD and NM models were found very close to actual values, but MSPE, RESC, ε and σ values were the lowest than actual. The highest R2 and θ values were found in these models. As the result of assessing of TMY, PY, PT, P, R2, MSPE, RESC, ε , σ , θ and DW statistics together, it can be said that the using AS, WD and new developed NM models make possible of predicting milk yields, close to actual values in Holstein cows at first lactation. Acknowledgments This research was funded in part by a grant from the University of Selcuk (BAP), The authors wish to thank the staff of KAR-YEM AŞ, Konya, TURKEY. References Ali, T.E. & Schaeffer, L.R. (1987). Accounting for covariance among across weeks of peak production test day milk yields in dairy cows. Can. J. Anim. Sci., 67: 637-644. Batra, T.R. (1986). Comparison of two mathematical models in fitting lactation curves for pureline and crossline dairy cows. Can. J. Anim. Sci., 66: 405-414. Bianchini, E.S. (1984). Estudo da curva de lactaçao de vacas da raça Gir. Doctoral Thesis. Faculta de Medicina de Ribeirao Preto, USP. Brody, S.A., Ragsdale, A.C. & Turner, C.W. (1923). The rate of decline of milk secretion with the advance of the period of lactation. J. Gen. Physiol., 5: 441-444. Chang, Y.M, Rekaya, R., Gionala, D. & Thomas, D.L. (2001). Genetic variation of lactation curves in dairy sheep: a Bayesian analysis of Wood’s function. Livest. Prod. Sci., 71: 241-251. Cobby, J.M. & Le Du, Y.L.P. (1978). On fitting curves to lactation data. Anim. Prod., 26: 127-133. Çilek, S. & Keskin, I. (2008). Comparison of six different mathematical models to the lactation curve of Simmental cows reared in Kazova State Farm. J. Anim. Vet. Adv., 7 (10): 1316-1319. Dalal, D.S., Malik, Z.S., Chhikara, B.S. & Ramesh, C. (2004). Prediction of lifetime milk production from early lactation traits in Hariana cattle. Indian J. Anim. Sci., 74:11. Dave, B.K. (1971). First lactation curve of Indian water buffalo. JNKVV Res. J. 5:93. Dědková, L. & Němcová, E. (2003). Factors affecting the shape of lactation curves of Holstein cows in the Czech Republic. Czech J. Anim. Sci., 48 (10): 395–402. Dhanoa, M.S. (1981). A note on an alternative form of the lactation model of Wood. Anim. Prod., 32: 349. Dijkstra, J., France, J., Dhonoa, M.S., Maas, J.A., Hanigan, M.D., Rook, A.J. & Beever, D.E. (1997). A model to describe growth patterns of the mammary gland during pregnancy and lactation. J. Dairy Sci., 60: 2340-2354. Dudouet, E. (1982). Courbe de latation thèorique de la chèvre et applications (theoretical lactation curve of the goat and its applications). Le Point Veterinare, 14: 53-61. 253 �Gengler, N. (1996). Persistency of lactation yields: a review. Proceedings of the Interbull Annual Meeting. Gemblouxs, Belgium, January 21-23. Bulletin No. 12. Department of Animal Breeding and Genetics, SLU, Uppsala, Sweden, 8796 pp. Gibson, T.A. & Grossman, M. (1990). Lactation curves in dairy goats: a review. Small Rumin. Res., 3: 383-396. Goodal, E.A. & Sprevak, D., 1985. A Bayesian estimation of the lactation curve of a dairy cow. Anim. Prod., 40: 189-193. Grossman, M. & Koops, W.J. (1988). Multiphasic analysis of lactation curves in dairy cattle. J. Dairy Sci., 71: 1598-1608. Grossman, M., Kuck, A.L. & Nortan, H.W. (1986). Lactation curves of purebred and crossbred dairy cattle. J. Dairy Sci., 69: 195-203. Guo, Z. & Swalve, H. H. (1995). Modeling of the lactation curve as a sub-model in the evaluation of test day records. Proceedings of the Interbull Annual Meeting. Prague, Czech Republic, Sep. 7-8 1. Bulletin No. 11. Department of Animal Breeding and Genetics, SLU, Uppsala, Sweden, 4 pp. Jenkins, T.G. & Ferrell, C.L. (1984). A note on lactation curves of crossbred cows. Anim. Prod., 39: 479-482. Keown, J.F. & Van Vleck L.D. (1973). Extending lactation records in progress to 305 day equivalent. J. Dairy Sci. 56: 10701079. Keskin, I. & Tozluca, A. (2004). Describing of different mathematical models for lactation curve and estimation of control interval in dairy cattle. Selcuk Univ. The J. of Agric. Fac., 18 (34): 11-19. Keskin, I., Dag, B. & Sariyel, V. (2009). Fitness of four different mathematical models to the lactation curve of Brown Swiss cows in Konya Province of Turkey. Can. J. Anim. Sci., 89: 195-199. Koçak, Ö. & Ekiz, B. (2006). Studies on factors affecting the milk yield and lactation curve of Holstein cows in intensive conditions. The Journal of The Faculty of Veterinary Medicine Istanbul University, 32 (2): 1-13. Landete-Castillejos, T. & Gallego, L. (2000). Technical Note: The ability of mathematical models to describe the shape of lactation curves. J. Anim. Sci., 78: 3010-3013. Masselin, S., Sauvant, D., Chapoutot, P. & Milan, D. (1987). Adjustment models for lactation curves. Ann. Zootechnie, 36: 171-206. Morant, S.V. & Gnanasakthy, A. (1989). A new approach to the mathematical formulation of lactation curves. Anim. Prod., 49: 151-162. Neal, H.D. & Thornley, J.H.M. (1983). The lactation curve in cattle: a mathematical model of the mammary gland. J. Agric. Sci., (Camb.) 101: 389-400. Nelder, J.A. (1966). Invense polynomials, a useful group of multi-factor response functions. Biometrics, 22: 128. TurkStat (2007). Livestock statistics. Turkish Statistical Institute. Ankara, Turkey. Olori, V.E, Brotherstone, S., Hill, W.G. & McGuirk, B.J. (1999). Fit of standard models of the lactation curve to weekly records of milk production of cows in a single herd. Livest. Prod. Sci., 58: 55-63. Papajcsik, I.A. & Bodero, J. (1988). Modelling lactation curves of Friesian cows in a subtropical climate. Anim. Prod., 47: 201-207. Ruiz, R., Oregui, L.M. & Herrero, M. (2000). Comparison of models for describing the lactation curve of Latxa sheep and an analysis of factors affecting milk yield. J. Dairy Sci., 83: 2709-2719. Sauvant, D. & Fehr, P. (1975). Classification des courbes de lactation et d’evolution de la composition du lait de la chèvre (classification od lactation curves and of the evolution of composition of goat milk). J. Recherche Ovine et Caprine (Paris), 2-4 Dec. 90-107 pp. Sikka, L.C. (1950). A study of lactations as affected by heredity and environment. J. Dairy Res., 17: 231-252. Statistica for Windows PC 5.0, (1995). Stat Soft, Inc. 2325 East 13th Street, U. S. A. Sturtevant, E. L. (1986). Influence of Distance From Calving on Milk Yield. Agri. Exper. Station Pep., 22-23, Geneva. Vargas, B., Koops, W.J., Herroro, M. & Van Arendonk, J.A.M. (2000). Modelling extended lactation of dairy cows. J. Dairy Sci., 83: 1371-1380. 254 �Wilmink, J.B.M. (1987). Adjustment of test day milk, fat and protein yield for age, season and stage of lactation. Livest. Prod. Sci., 16: 335-348. Wood, P.D.P. (1967). Algebraic model of lactation curve in cattle. Nature, Lond. 216: 164-165. 255 � Dublin Core The Dublin Core metadata element set is common to all Omeka records, including items, files, and collections. For more information see, http://dublincore.org/documents/dces/. Extent The size or duration of the resource. 435 Title A name given to the resource Comparison of Eleven Mathematical Models for describing the first Lactation Curve of Holstein Cattle in Turkey Author Author Keskin, İsmail Memmedova, Nazire İlhan, Fatma Dağ, Birol Mikailsoy, Fariz Abstract A summary of the resource. In this study, eleven standard lactation curve models (Incomplete Gamma (WD), Quadratic (Q), Cubic (C), Linear Hyperbolic Function (LH), Inverse Polynomial Function (IP), Mixed Log (MIL), Exponential (WIL), Dhanoa (DH), Cobby and Le Du (CD), Polynomial Regression (AS) and New Model (NM)) were used to predict a typical dairy cow lactation derived as the average daily milk yield of 105 complete first lactations of Holstein- Friesian cows in one herd. Milk yield controls were made daily in this farm. Total milk yield (TMY) was calculated from observation of daily milk yield. TMY was also predicted by using 11 different models. The total milk yields predicted by the models were very close to each other and the differences between observation of TMY were not found statistically significant (P>0.05). The models were found to be adequate for estimation of milk yield. Determination coefficients (R2) of the models ranged from 67.15 % to 86.68 %. In comparison of the models, the TMY, peak yield (PY), peak time (PT), persistency (P), mean square prediction error (MSPE), approximation error ( ε ), reliability criterion for estimating trustworthiness of the determination coefficient (σ ), standard error ( θ ) and Durbin-Watson (DW) values were evaluated together. Using the AS, WD and new developed NM models accurately predicted the milk yield of Holstein cows. Date A point or period of time associated with an event in the lifecycle of the resource 2010-06 Keywords Keywords. Conference or Workshop Item PeerReviewed Q Science (General) https://omeka.ibu.edu.ba/files/original/45f095bd0421f6b027d9bd2d4d6effc8.pdf b5b7a252e2972987e8be2e1fe7d2472c PDF Text Text 2nd International Symposium on Sustainable Development, June 8-9 2010, Sarajevo Comparison of Islamic, Traditional and Alternative Utility Theories Sümeyye DEMĠRSOY International University of Sarajevo Bosnia and Herzegovina Mehmet CAN International University of Sarajevo Bosnia and Herzegovina mcan@ius.edu.ba Abstract: Decision making under uncertainty is always trying to be explained by utility theory. And utility theory‘s roots rely on moral philosophy. Moral philosophy is concerning concepts about good and bad, right and wrong, virtue, justice, etc. It can be say that utilitarianism, which is a field of moral philosophy, is more directly about utility theory. Throughout the human history, from Prophet Abraham to Greek philosophers; Socrates, Aristotle, Epicurus, to Islam scholars al Kindi, al-Farabi, al-Razi, Ibn-i Sina, Ibni Rushd, Ibn-i Haldun, all discussed about ethics and utility concept. 1. Introduction Although utility is an economic term which measures the satisfaction or desirability in terms of the consumption of goods and services, its roots rely on moral philosophy which deals with the concepts about good and bad, right and wrong, virtue, justice, and happiness. On the other hand, concepts of good and bad, right and wrong, virtue, justice, and happiness was constituents of ethichs throughout history of humanity. Human concepts like good and bad, right and wrong, virtue, justice, and happiness were the concern of human civilizations through millenniums. Historically the foundations of human ethics are laid by divine revaliations through prophets. Muslims identify the prophets of Islam as those humans chosen by Allah to teach mankind. Humans may rely on revelation or tradition to identify prophets. Each prophet brought the same basic ideas of ethics. They brought the belief in a single God and in the avoidance of idolatry and sin. Muslims regard Adam as the first prophet and Muhammad as the last. Islamic theology recognises as many as 124,000 prophets. The Qur'an identifies 25 prophets by name, starting with Adam and ending with Muhammad. Five of them, Rasuls, receive the highest reverence for their perseverance, Ibrahim (Abraham), Moosa (Moses), Dawud (David), Isa (Jesus), and Muhammad. 1.1 Prophet Ibrahim (Abraham) Ibrahim was born in a house of idolaters, in the kingdom of Babylon. He announced to his people: O my people I turned my face towards Him Who created the heavens and the earth, and never shall I give partners to Allah. He has the power to make the stars rise and set. 1.2 Prophet Musa (Moses) Musa the son of Imran, was born in Egypt in which at that time the kings were known as Fir‘awns. First statement about working ethics in Torah is in Genesis part. ―In the sweat of thy face shalt thou eat bread, till thou return unto the ground; for out of it wast thou taken: for dust thou art, and unto dust shalt thou return.‖ (Torah, Genesis 3/19) In Torah it is suggested to respect neighbour rights and to behave in good way to them: ―When thou dost lend thy brother any thing, thou shalt not go into his house to fetch his pledge. Thou shalt stand abroad, and the man to whom thou dost lend shall bring out the pledge abroad unto thee. And if the man be poor, thou shalt not sleep with his pledge: In any case thou shalt deliver him the pledge again when the sun goeth down, that he may sleep in his 476 �2nd International Symposium on Sustainable Development, June 8-9 2010, Sarajevo own raiment, and bless thee: and it shall be righteousness unto thee before the Lord thy God."(Torah, Deureronomy 24/10-13) Another statement about poor is as following and in this statement Torah mentioning about living together with brothers: ―And if thy brother be waxen poor, and fallen in decay with thee; then thou shalt relieve him: yea, though he be a stranger, or a sojourner; that he may live with thee. Take thou no usury of him, or increase: but fear thy God; that thy brother may live with thee.‖ (Torah, Levitucus 25/35-36) ―If thou lend money to any of my people that is poor by thee, thou shalt not be to him as an usurer, neither shalt thou lay upon him usury. If thou at all take thy neighbor's raiment to pledge, thou shalt deliver it unto him by that the sun goeth down: For that is his covering only, it is his raiment for his skin: wherein shall he sleep? and it shall come to pass, when he crieth unto me, that I will hear; for I am gracious.‖ (Torah, Exodus, 22/25-27) ―And when ye reap the harvest of your land, thou shalt not wholly reap the corners of thy field, neither shalt thou gather the gleanings of thy harvest. And thou shalt not glean thy vineyard, neither shalt thou gather every grape of thy vineyard; thou shalt leave them for the poor and stranger: I am the LORD your God.‖ (Torah, Levitucus 19/910) ―And if thy brother be waxen poor, and fallen in decay with thee; then thou shalt relieve him: yea, though he be a stranger, or a sojourner; that he may live with thee.‖ (Torah, Levitucus 25/35) 1.3 Prophet Dawud (David) Dawud (David) was not only an illustrious Prophet of the Israelites but he was also their king. The Holy Qur'an affirms: "And Allah gave him the kingdom! and wisdom and taught him of that which He willed." (2: 251). And it was said unto him: ― 0' Dawud!To! We have set you as a vicegerent in the earth, therefore, judge aright between mankind and follow not desire." (38:26) He lived in Bait-ul-Lahm which was situated at the distance of ten miles from Jerusalem. He prayed: "Our Lord! Pour out constancy on us and make our steps firm and help us against those who are disbelievers." (2:249) Allah revealed the Zabur (Book of Psalms) to Prophet Dawud. It contains lessons for the guidance of his people. 1.4 Buddha The evidence of the early texts suggests that the Buddha was born in a community that was on the periphery, both geographically and culturally, of fifth century BCE northeast India. This community seems to have had two categories of people, masters and servants. The Four Noble Truths of Budhism: 1. 2. 3. 4. Life as we know it ultimately is or leads to suffering/uneasiness (dukkha) in one way or another. Suffering is caused by craving. This is often expressed as a deluded clinging to a certain sense of existence, to selfhood, or to the things or phenomena that we consider the cause of happiness or unhappiness. Craving also has its negative aspect, i.e. one craves that a certain state of affairs not exist. Suffering ends when craving ends. This is achieved by eliminating delusion, thereby reaching a liberated state of Enlightenment (bodhi); Reaching this liberated state is achieved by following the path laid out by the Buddha. 2. The Greek Philosophers For ancient Greek philosophers the question ‗how should I live?‘ took a fundamentally prudential or selfregarding form. It amounted for them an inquiry searching how a man could secure his own happiness, fulfilment or perfection. Benevolence, altruism, philanthropy, a concern for happiness of others occupied a secondary position in their ethical recommendations. It is conceived as a condition of self-realisation of the individual. In general Greek philosophers, Plato and Aristotle in particular, found a place for restricted benevolence by emphasising the role of friendship in a fully satisfying life. Aristo somewhat made a disdainful liberality part of his conception of the ethically ideal or ‗magnanimous‘ man. It can be said that utilitarianism, which is a field of moral philosophy, is more directly about utility theory. Utilitarianism can be understood as a movement for legal, political and social reform that flourished in the nineteenth century. It can also be understood as the ideology of that movement. It is also a general ethical theory. As a theory of 477 �2nd International Symposium on Sustainable Development, June 8-9 2010, Sarajevo ethics, it provides a criterion for distinguishing between the right and wrong action, an account of nature of the moral judgements that characterise action as right and wrong. Utilitarianism can be expressed as the combination of two principles: (i) the consequentialist principle that the rightness, or wrongness of an action is determined by the goodness, or badness of the results that flow from it and (ii) the hedonist principle that the only thing that is good in itself is pleasure and the only thing bad in itself is pain. The doctrine can be expressed in the form of a single principle, the greatest happiness principle: the rigthness of an action is determined by its contribution to the happiness of everyone affected by it. (Quinton, 1973). Greek moralists unreflectively assumed a consequentialist position in developing their more or less prudential life-styles. The only way in which they conceived it to be possible to justify a type of conduct was by reference to the results to which it gave rise (Quinton, 1973). 2.1 Socrates (BC 470-399) Socrates' ethical intellectualism has an eudaemological character in which he asserted that the highest good for any human being is happiness. According to Socrates, people‘s true happiness is promoted by doing what is right. When people‘s true utility is served (tending own soul), people are achieving happiness. Happiness is evident from the long-term effect on the soul. According to him, whatever action a man chooses is motivated by his desire for happiness. All the concepts like knowledge, virtue, and wisdom are the same and man chooses an action according to what he thinks will bring him the greatest happiness. Therefore the more a man knows, the greater his ability to reason out the correct choice and the greater his ability to choose those actions which truly bring happiness to him. According to Socrates, to answer the question ‗what is happiness‘, the first question should be asked to an individual by himself: If all his needs were completed, or he had enough power to do anything, would he really be happy? When individual observe and attract attention to him, even all these needs mentioned above are completed, he sees that he is not happy enough. On the contrary he sees that he witnessed many disappointments, only when he is in harmony with himself he will really be happy. People who are not in harmony with themselves never be happy properly. Socrates believes that psychic harmony is the greatest good, and that the result of it is moral behavior. He also believes that if you have a healthy body and soul then you are in psychic harmony with yourself. In the ideal soul, the reasoning part and the feelings (desires for honour) rule over the appetitive part (desires for wealth, food, etc). A properly ordered soul experiences a sense of well-being or psychological health. Thus, psychological health is something distinct from psychological stability since it depends on psychic harmony. According to Socrates, immoral behavior is a result of an unbalanced personality and leads people to irrational behavior. Psychic harmony is a psychological condition and makes someone moral and this harmony has no motivation. Moral behavior comes from people‘s own beliefs and desires. If one is bad or unjust in the social sense it is because of his sensuality, greed, or vanity. According to Socrates where there is psychic harmony, the motives for injustice in the social sense will be eliminated. Socrates did not surpass the prejudice of Greek intellectualism in ethics. It is enough to know virtue in order to be virtuous. Everyone wishes to be happy and if someone does not attain happiness, it is because he does not know the way that leads to happiness. Consequently, so-called evil men are in reality only ignorant. Thus, vice is synonymous with ignorance, so knowledge of the good is synonymous with virtue. That is why Socrates, who intended to form a virtuous youth, restricted his teaching to the search for moral concepts. The foundations of Socratic system of ethics can be summarised as; i. a choice is rational if and only if it is a choice of what is best for the agent ii. something is good for an agent if and only if it is morally right. The cornerstone of Socratic ethics is the self-interested concern for happiness, that is, one's own good. But this concern requires that we act in accordance with what is just and noble, that is the moral good. The identity of one's own good with the moral good is the basis for specific Socratic foundations. Socrates shows in the book Gorgias that one's own good coincides almost completely with the moral good, since the utility of nonmoral goods, for example, wealth, depends on the possession of moral excellence. According to Socrates, the content of goodness/benevolence covers utility and pleasure. That is why some of the historians of ethics assume that Socrates is Epicurean, so he is utilitarian. However, pleasure, which Socrates implies, is not harmfull to intellect and soul. Here intellect has a role of being informative and determinant. Thus Socrates‘ pleasure is under control of intellect. And also Socrates‘ utility view is not individualistic, but public. 478 �2nd International Symposium on Sustainable Development, June 8-9 2010, Sarajevo 2.2 School of Cyrene – Aristippus (BC 435-366) Cyrene school of philosophy is a kind of discipline of Socrates, founded in the 5th century by Aristippus. He was a hedonist and urging the pre-eminent claims of bodily pleasure as an end. This philosopher explains his thought, which he starts with a question ‗what is a happy life‘, by answering ‗life, which has as much as possible pleasure and as little as possible suffering, is a happy life‘. To make life happy, someone should include as much as possible pleasure and as little as possible suffering to his/her life. For to do this, he/she should decrease needs, so it will be adjusted to life with limited needs. So by doing with limited needs, it can be benefited from all pleasures of life. Aristippus teaches us kind of a life art that should be lived wisely. And this art can be gained by giving up desires which make people slaves of themselves. Thus, school of Cyrene derived hedonism from Socrates‘ eudaimonism. Later on, school of Cyrene‘s this hedonism concept was accepted by Epicurus (Aster, 2005) 2.3 School of Cynicism – Antisthenes (BC 455-365) Antisthenes, who outlined the themes of the school, is the first philosopher of school of cynicism. He had been a pupil of Socrates in the late 5th century BC. Antisthenes has severe discussions to Aristippus‘ pleasure assumption. According to Antisthenes, people should look for real happiness in desire of freedom and liberty inside. People, who reach to real happiness, know how to stay insensitive and disregarding in relation to either pleasure or suffering. This staying indifferent in relation to pleasure or suffering brings freedom from inside. 2.4 Plato (BC 427-347) Plato, founder of Academy in Athens, is a classical Greek philosopher. His mentor was Socrates and his student was Aristotle. According to Plato, nothing is fine without moderation, and the thing which is pleasant becomes as pleasant with this moderation. He also mentiones about moderation in between organic pleasures and intellectual benevolence. Unlike Aristotle‘s developed concept of happiness (will be mentioned below), Plato‘s happiness concept is more obscure. According to Plato, the Good is the source of intelligibility. He asserted that the highest goal in all of education is knowledge of the Good. According to him, human beings aim at the good, nobody voluntarily chooses evil. Plato thinks that the masses are incapable of grasping the truth. He illustrated this in his the allegory of the cave where Plato suggests that the masses cannot see the truth directly but they are satisfied with an illusion of reality. According to Plato, the good is the source of intelligibility and happiness is the attainment of intelligibility. Thus, Plato describes his happiness concepts as the goal of life. There is an important difference between Socratics and Plato. Both Aristippus and Antisthenes are individualists. For both of them, origin is individual. Both of them don‘t deal with super-individual fact as state, history, society. If people want to be really happy, they should be on their own, they should not depend on other people. In this point, Plato has disagreement with his other two schoolmates. According to Plato, people never be on their own but always live with other people. If we isolate an individual from the society he lives in, we exclude him from his own resources. Moreover, the institution called ‗state‘ is like a human. What condition social entire is in, individual lives in the same condition. Thus to understand human being, it should be looked into the state which he lives. Therefore Plato rejects Socratics‘ (Aristippus and Antisthenes) thoughts. Although they seem like the followers of Socrates, Plato thinks that they are not thinking in similar perspective. According to Plato, contrary to Aristippus and Antisthenes‘ individualistic tendency, Socrates did not only show his skillfulness about living, but he also became the first example of moral principle which took socialization as a base (Aster, 2005). 2.5 Aristotle (BC 384-322) Aristotle is a Greek philosopher who is student of Plato and teacher of Alexander the Great. His teaching about virtue and ethics is set forth in his Nicomachean Ethics. Greek philosophy accepts that human life‘s main purpose is ‗happiness‘. However they differentiate from each other about ‗what is happiness‘. Aristotelian method is different from others because, according to him, each existence has 479 �2nd International Symposium on Sustainable Development, June 8-9 2010, Sarajevo its own activity. Knowing specific activity of human shows us that what kind of an objective this human wants to reach. So it will be learned what is the ‗real happiness‘. According to Aristotle, happiness (eudaimonia) is an activity of the soul in accordance with virtue and it is the highest of all goods. Happiness is the first principle and cause of all goods and it is a self-sufficient activity always chosen for itself. Unlike other virtues, happiness is the only thing chosen only for itself and not for the sake of other things. Aristotle believes that amusement is not self sufficient like happiness and he distinguished between the life of amusement and the happy life. Aristotle proposes two possible paths to happiness as a life of virtuous activity and a life of theoria and asks which of the two is the best. The former path defines happiness as a practical vitue and in need of external goods. And the latter path theoria is contemplation of eternal truths for an entire lifetime which is the highest activity of reason. Aristotle chooses theoria. Becuase the life of practical virtue achieves happiness in a lesser sense because of the necessity of material goods in this life and the life of the theoria limits the need for material goods and perfect happiness (eudaimonia) is enjoyed by the Gods. According to Aristotle, human being is rational as his nature. If a rational man behaves rationally and moderate, he will be behaved parallelly to his nature. ‗Thinking‘ and ‗knowing‘ are human‘s highest activities. But what can be ratio/intellect‘s activity in practical life? Each of the virtues is a state of being that naturally seeks its mean. All virtues are in between of two extremes, virtues are means of the two extremes. This mean is not mathematical but it is a border which can be found by ratio/intellect. The virtuous habit of action is always an intermediate state between the opposed vices of excess and deficiency. For example; with respect to the enjoyment of pleasures, temperance (sophrosúnê) is a mean between the excess of intemperance and the deficiency of insensibility. Greek thought always searches for harmony and moderation, dislikes extremes. Aristotle is an opposer to all kind of extremes, too. Thus, Aristotle is against the thoughts that reject possessions like Cynics. Aristotle emphasizes on possessions‘ relative values. By using them moderately, they can bring even sense and value. One another thought of Aristotle is not to ignore desires and urges totally, they should not be ignored totally, but to keep them down by ratio. According to Aristotle, there is no direct connection between good/benevolence and pleasure. These three points should be taken into account: First pleasure is not main principle for a moral life, second pleasure occurs as a result of an action which targets virtue. And third virtue is in the action which results in pleasure. 2.6 Epicurus (BC 341-271) Two more schools joined to the list of schools of philosophy in about BC 300s. One is school of Stoa and the other one is Epicureanism that took its name from the founder Epicurus. These two schools have contradictory opinions about life and knowledge. As it is mentioned in former parts of this work, school of Cynicism and Cyrene have contradictory opinions about ethics, too. According to Cynics, it is important to exert perfect dominance on desires. After Cynics, we see some similar thoughts are repeated by Stoa. It is virtuous to exert dominance on excitements, desires because virtue gives us staying apathetic to life and death. School of Cyrene perceives life‘s real purpose as catching pleasure and escaping from pain. Kind of a dissidence, like between Cynics and Cyrenes, later on was seen in between Stoa and Epicureanism. These two schools, Stoa and Epicureanism, sustained their existence by keeping alive dispute among them. However at the same time they have some similar or common opinions. The first point they agree is that human being is subject of philosophy. First they draw portraits of ‗superman‘. However all these ‗superman‘ concepts were interpreted differently. For Stoa school, superman is man who beats demands and desires, knows to disregard to both life and death. They show apathy (lack of interest or concern) as a purpose to human. On the other hand Epicureanism finds its purpose on ataraxie (freedom from worry). However it can be noticed that there is no big difference between ataraxie and apathy. According to Stoa school, first principle is people‘s need to understand that they are organs of the unity called world. Second principle is people‘s need to know their own stand in the world and so to adopt themselves to destiny which is chosen for themselves. However Epicurus thinks that world processes with respect to blind and spontaneous necessity. If people‘s destinty is determined by coincidences which can not be seen before, then they can have interest in products which are their own will-power. Thus, people will stay disregard in respect to life and death and by behaving rationally they will know how to distinguish things which give happiness. Epicurus adopted the principle to attain pleasure and to escape from pain in ethics. But human should do this wisely. He should avoid from intense pleasures which bring pain at last. Human should not interest in anything more than its necessity because extremes cause to pain. People should know to keep away from temporary and 480 �2nd International Symposium on Sustainable Development, June 8-9 2010, Sarajevo specious values such as fine and glory. Temporary values always direct people to more, and this more never ends, that is why people always stay in unrest. Thus, people should be interested in ―moral pleasures‖ which won‘t cause to dissatisfaction. According to Epicurus, for to be happy, it is necessary to live moderate, to tend pleasures which are moral, and to behave in accordance with all these. Epicurus established his school in Athens in 306 BC. In the surviving writings of Epicurus, there is little of direct relevance to the connection between utility and justice. However, virtue including justice, was not intended to limit pleasure. According to Bailey (Bailey 1928) Epicureanism is ‗a system of uncompromising egoistic hedonism‘. Scarre (Scarre 1994) put it, ‗just as the Epicurean community practiced economic self-sufficiency within the walls of its garden, the Epicurean man cultivates an inner self-sufficiency, a contentment in his own physical and mental states and a suppression of unnecessary desires‘. The only perfect pleasure was a condition of ataraxie where one lives quietly in bodily health and with little physical and psychological distress. (Rosen, 2003). Although none of the pleasure is bad, some (involving less pain) are purer than others. Epicurus made connection between pleasure and health, pain and disease. All pleasures were good in the sense that health was good, even though some pleasures were mixed. If health was good, with disease of body or soul it becomes greatest evil. In Epicureanism the greatest pleasure was defined by the removal of all pain, and hence the Epicurean lived quietly and peacefully in the real or metaphorical Garden (Rosen, 2003). The important virtue for Epicurus was prudence, and a considerable emphasis was placed on the egoistic pleasures connected with friendship. On the other hand, little attention was given to social values and instincts. When it comes to ‗justice‘ in Epicurus‘ system, it can be said that ‗justice‘ means achieving security from the attacks of other people. Epicurus‘ state on a conception of justice is that; a pledge of mutual advantage to restain men from harming one another and save them from being harmed (Epicurus,1926). At another point, Epicurus wrote about justice in terms of being of advantage in the requirements of men‘s dealings with one another (Epicurus, 1926). Epicurus used here a Greek phrase, ‗sumpherei en tais chreiais‘, for ‗advantage in the requirements‘. Rosen claims that Greek noun, ‗chreia‘, also possessed a range of meanings and might be translated as ‗need‘, ‗use‘ or ‗utility‘. For all societies in which to make compacts not to harm one another is not possible, nothing was either just or unjust. Although justice was applied potentially to all requiring, where contracts for one should not harm others, such a justice might be applied differently in different societies and under different circumstances. Where a law, which was previously considered, just, had no longer had usefulness or secured advantage, there was no longer just (Epicurus, 1926). According to Alberti (Alberti 1995) ‗justice is the realization of utility by means of a contract‘. The emphasis on utility allows for the separation of law from justice by rejecting the view found in Plato and Aristotle, that all law is just. And emphasis on utility leads to a notion of justice which is different from nomos (legal justice) and physis (natural justice) (Rosen, 2003). Justice was an invention of the wise for their own good. Epicurus summed up matters with brutal directness and claimed that the laws exist for the sake of the wise, not that they may not do wrong, but that they may not suffer it (Bailey,1928). Law and justice were matters of convenience which the wise person devised and approved. Epicurus had no reason to make justice a positive part of the human condition except enabling people to obtain ‗peace of soul‘ (DeWitt, 1954). ―It represented a painful burden, and in its application as punishment justice could be extremely painful. All that could recommend it was its utility to the wise. Other members of society might have less invested in justice, as they were not cultivating their gardens as were Epicureans, and might well gain less from rules concerning not harming others. However, so long as they accepted the compact, they would be assisting themselves, as well as not harming the wise in society.‖(Rosen, 2003). 3. Medieval Times Islam Philosophers In the seventh century, translation movements from Greek to Arabic language started and in the time of Caliph al-Mansur this movements reached its peak. The study of Islamic ethics began to take shape in the third century of Islam‘s emerge, with the influences from Greek ethics including Stoicism, Platonism and Aristotelianism. Al-Kindi, the first philosopher of Islam, influenced by Socrates and Diogenes the Cynic as seen in his ethical writings. Other influences can be seen in the work of Platonists such as Abu Bakr al-Razi or Neoplatonists such as al-Farabi, Aristotelian influences can be seen in the works of al-Farabi, Ibn Sina and Ibn Rushd. 481 �2nd International Symposium on Sustainable Development, June 8-9 2010, Sarajevo 3.1 Al-Kindi (d.873) Abu Yusuf Ya‘qub ibn Ishaq al-Kindi (d.873) was the first philosopher of Islam and also the first author on philosophical ethics. In Baghdad, al-Kindi was involved in the scientific movement of the translation of Greek texts into Arabic. His starting point was Greek philosophy and he is reported by the classical bibliographers that he has a number of ethical treatises reflecting an interest in Socratic and Cynic thought. In al-Kindi‘s writings, the personalities of Socrates and Diogenes the Cynic are united and both emerge as ideal instances of virtue and asceticism (Fakhry, 1998). Moreover, the Stoic idea of apatheia (freedom from passion) and the indifference to the vicissitudes of fortune are set out in fluent terms. According to al-Kindi, the antidote of pain is to consider that pain results either from our actions of from doing the actions of others. In the former case, it is individual‘s duty to avoid from doing which is the cause of pain. In the latter case, averting the pain is either in our power or it is not. If it is in our power then we certainly ought to avert it, if it is not, we should not suffer at the prospect of injury with the hope that it might somehow be turned away. The suggestion, to avoid material possessions as temporary acquisitions, reflects the influences of the Stoic philosophers. 3.2 Abu Bakr al-Razi (d. 925) Another philosopher Abu Bakr al-Razi (d. 925), who is influenced by Plato, refers to Plato as ‗the master of the philosophers and their leader‘ and to Socrates as ‗the ascetic and spritual‘ sage in his book al-Tibb al-Ruhani (The Spiritual Physic). A Socratic-Platonic theme which takes place in al-Razi‘s writings is the foolishness of the hedonistic life which turns man into a slave. People‘s many of pleasures are temporary and unattainable and people are attacked by anxiety or pain. But according to al-Razi, true philosopher will not succumb to pain, because philosopher understands that nothing is permanent in this world. And that whatever cannot be turned away should be ignored, since it is the product of passion and not of reason. al-Razi says in his book al-Falsafiya that: 'For reason summons us only to what is susceptible of bringing about profit sooner or later; grief does not bring any advantage... That is why the perfectly rational man will only follow the summons of reason ... and will never follow the summons of passion or allow himself to be led by it or get close to it. Like Socrates and Plato, al-Razi believes that the soul, leaving the body, will return to its original residence in the intelligible world, after passing through an endless cycle of purifications. Death is a logical consequence of our being human and essential part of the definition of man. However, al-Razi adds another argument which derives from Epicurus that death is the deprivation of sensation and with his death man will be stripped of the sensations of pleasure and pain. Thus this is a better condition than living in pain. That is why 'according to the judgment of reason the condition of death is better than the condition of life' (Rasa‘il al-Razi al-Falsafiya). 3.3 Al-Farabi (d. 951) Abu Nasr al-Farabi (d. 951) was known as the ―second master‖ (muallim-i sani) amongst his peers, Aristotle being the first (muallim-i evvel). Al-Farabi was the first systematic writer on philosophical questions in Islam. He also contributed to ethical discussions and wrote a commentary on some parts of the Nicomachean Ethics which is translated into Arabic by Ishaq bin Hunayn. Al-Farabi follows Aristotle in ethics like dividing the virtues into moral and intellectual (Fakhry, 1998). According to him, moral virtues are perfections of the lustful part of the soul whereas perfections of the intellectual part are practical reasoning, good judgement, wisdom and sound understanding. Al-Farabi also follows Aristotle‘s arguments about justice which consists in the equitable distribution of ‗common goods‘ in the city or in the state. Every member of city or state is entitled to share this ‗common goods‘ such as security, wealth, dignity, public office, etc. Al-Farabi deflects from Aristotle and other Greek phlosophers believing in the life hereafter in Qor‘anic way. Accoring to al-Farabi, nations and citizens of cities attain happiness, worldly in here and supreme in the life hereafter, when four human needs are met. These are; theoretical virtues, deliberative virtues, moral virtues and practical arts. Worldly happiness is necessary for the attainment of supreme happiness in the hereafter. According to him, happiness is the absolute good and achieving happiness is the purpose of life. Whenever the soul of the person reaches perfection, then happiness is achieved. According to al-Farabi, if individual‘s desire for happiness is weak and he/she has other purposes in his/her life, then the result will be evil. 482 �2nd International Symposium on Sustainable Development, June 8-9 2010, Sarajevo Theoretical virtues, the first one of the four themes mentioned above, consist of the sciences. The purpose is to have understanding of all the beings on these sciences. Deliberative virtues are voluntary intelligibles that vary across time and place such as events occuring accidentally or willingly, such as disasters or war. An individual cannot possess deliberative virtue without possessing moral virtue. A person has to have virtuous moral character who wishes the good for himself/herself or for others. And according to al-Farabi, theoretical virtues, deliberative virtues, moral virtues and practical arts are all inseparable. The his famous work al-Madina al-Fadila (virtuous city), in which people cooperate and help each other for the purpose of attaining happiness, al-Farabi reflects again his devotion to Islamic utility perception. To him, on the other hand, non-virtuous city is the city whose people do not know happiness. Al-Farabi discussed these two cities‘ souls of the citizens and he believed that the souls of the citizens of the virtuous city are immortal. However, the souls of the citizens of the ignorant city are mortal and their destiny is to suffer. As a result, al-Farabi believed that political association should be directed towards the attainment of happiness. Generally, it can be said that al-Farabi was greatly influenced by both Aristotle and Plato in his philosophy and his concept of happiness particularly. On the other hand, his thought was framed by Islam. He selected portions from each of these three different influences to form a complete description of happiness. So his concept of happiness is a product of his understanding of Greek philosophy and Islam. With his concept of happiness, he combines Plato‘s concept of the good, Aristotle‘s concept of eudaimonia (happiness) and Islamic concept of Jihad Al nafs (struggle of the soul). Plato and Aristotle‘s concepts were given above. On the other hand, the Islamic concept of Jihad Al nafs means the struggle of the soul. According to Islam, God created man to achieve bliss (happiness) in the next life through a clearly defined struggle in this life called Jihad. By the Quranic definition, ―And whosoever strives (jahada), strives (yujahidu) only for himself. Surely Allah is self sufficient, above need of His creatures.‖ (Quran 29:6). The person who struggles with turning his inner self into a new way of living that understands the true reality where material is only a small portion. 3.4 Ibn Sina (Avicenna) (d. 980) Ibn Sina (Avicenna) (d. 980) is one of the foremost philosophers in the Medieval Hellenistic Islamic tradition and one of the most important practitioners of philosophy. He exercised a strong influence over the other Islamic philosophers and medieval Europe as well. Al-Farabi‘s successor Ibn Sina is the author of a very short tract on ethics and he follows closely the Platonic model in psychology. Ibn Sina speaks about the laws which are needed to be laid down as the moral habits (akhlaq) and traits (adat) which lead to justice. He divides the soul into different parts like rational, irascible, and concupiscent which correspond to the virtues of wisdom, courage and temperance respectively. Finally justice is the ‗summation‘ of all these three. According to Ibn Sina, enforcement of justice within the state (with the existence of caliph) is necessary as the sovereign of the world and God‘s vicegerent on earth. The virtues of temperance, courage, and wisdom are for the well-being of human beings in this world. They can be followed adequately without theoretical wisdom. Ibn Sina presents theoretical wisdom as being so important that one can attain happiness only by acquiring it as well as these three virtues, all of which add up to justice. Ibn Sina distinguishes himself from Farabi by insisting on the possibility of acquiring temperance, courage, and practical wisdom-or justice-without possessing theoretical wisdom. In other words, unlike Farabi, Ibn Sina does not perceive all the virtues to be intellectual or to be grounded in sound intellectual understanding. Make separation between the practical virtue and the theoretical virtue does not fully account for Ibn Sina's moral teaching. From what appears in his treatises that moral habits are directed to the liberation of the soul from the body. Thus they serve the ultimate goal of theoretical virtue, that is the soul achieving a free perception of God and the divine intelligences. It is not clear, however, how the moral habits lead to justice. The only explanation that comes to mind is that insofar as some human beings center their thoughts and activities on other worldly concerns. Ibn Sina differs here from al-Farabi too since he starts with the basic human needs and ascends from them to the larger issue of law- giving and providing for justice. On the other hand, Al-Farabi begins by thinking about ultimate human happiness. Ultimate human happiness is about the highest ends of human beings rather than their humblest beginnings or it is about their noble concerns rather than about their basic needs. 483 �2nd International Symposium on Sustainable Development, June 8-9 2010, Sarajevo 3.5 Ibn Rushd (Averroes) (d. 1198) Ibn Rushd (Averroes) (d. 1198) is regarded as one of the important Islamic philosophers. He set out to integrate Aristotelian philosophy with the Islamic thought in the twelfth-century of Islamic Spain. He produced commentaries on Aristotle‘s Nicomachean Ethics and also Plato‘s Republic which is relevant to his ethical theory as well. According to Ibn Rushd, the principal virtues correspond to the perfection of the three parts of the soul are the rational, the irascible, and the concupiscent. Then he describes justice along Platonic lines as the ‗harmony‘ of the three corresponding virtues of wisdom, courage and temperance. As Aristotle stated in the Nicomachean Ethics, it has two subdivisions which are common or universal, corresponding to ‗perfect virtue‘, and particular. However, Ibn Rushd does not identify happiness with the contemplative life, but rather with conjunction (ittisal) to the active intellect, which the Muslim Neoplatonists had regarded as man‘s ultimate goal. In Muslim thought, everything they need to know about moral behaviour is encapsulated in Islam. However, Ibn Rushd argued that a distinction should be drawn between moral notions and divine commands and here he follows an Aristotelian approach (Leaman). According to Ibn Rushd, the answer of question ―what is the purpose of a human being?‖ is that; one of the ultimate aims is to be happy and to avoid actions which lead to unhappiness. Moral virtue leads to happiness. If people do what they should do in accordance with their nature, people will be able to achieve happiness. This happiness may be interpreted as a mixture of social and religious activities or as an entirely intellectual ideal. However, neither religion nor philosophy would approve of entirely intellectual ideal as the ultimate aim for the majority of the community. It is possible for someone that he/she would try to live apart from the community with concentrating entirely on intellectual pursuits, but this way of living is inferior to a life in which there is a concentration upon intellectual thought but combined with integration within the practices of a particular society. Ibn Rushd, inspite of working within an Islamic context, does not identify happiness and misery with some aspect of the afterlife since he was unable to accept the traditional view of the afterlife. Here Oliver Leaman, who has many works about Ibn Rushd, claims that without religious imagery, ordinary believers may find it difficult to understand that our moral actions affect not only ourselves but the happiness of the whole community, not just at a particular time or in a particular place but as a species. With our bad behaviours, we damage our own chances of human flourishing, and this damage affects our personal opportunities for achieving happiness and maturing as people. It is also resulting to the weakening of society. According to Leaman, while it is possibly true that the misery of evil-doing may not follow us personally after our death, it may well follow the community. The notion of an afterlife points to the wider terms of reference in which moral action has life. 3.6 Ibn Khaldun Another philosopher Ibn Khaldun, who lived in 14th century, centred his economic ideas generally around the ideas of justice, hardwork, cooperation, moderation and fairness. He emphasises Al-adl (justice) as the bedrock of the economy, and lack of justice leads to the breakdown of the state. Some of his writings may appear as secular. For instance; ―Civilisation and its well-being as well as business prosperity depends on production and people‘s efforts in all directions in their own interest and profit‖ (Muqaddimah, Volume 2). However, Ibn Khaldun insisted that man must avoid from evils, must improve himself, and must give preference to matters of the next world against this world (Muqaddimah, Volume 1). According to Ibn Khaldun, extravagance and luxurious living lead to the destruction of the state. ―Sedentary people are much concerned with all kinds of pleasure. They are accustomed to luxury and success in worldly occupations and indulgence in worldly desires. Therefore, their souls are adored with all kinds blameworthy and evil qualities‖(Muqaddimah, Volume 1, 225). Another issue Ibn Khaldun emphasises is cooperation. He says that ―the power of the individual human being is not sufficient for him to obtain the food he needs through cooperation, the needs of a number of persons, many times greater than their own number can be satisfied‖ (Muqaddimah, Volume 1, 69) ( Ibn Haldun, 1977). Above the economic philosophy of Ibn Khaldun was mentioned shortly, but, his Muqaddimah covers a large number of other areas about economy like money, value, market, population, growth, international trade, etc. When dealing with micro or macro economic issues, he demonstrated perfect competence in generating theories. In his work, Ibn Khaldun synthesises the ideas learnt from the Qur‘an and Sunnah, and from other sources which were converted into powerful theories. 484 �2nd International Symposium on Sustainable Development, June 8-9 2010, Sarajevo 4. Conclusion Human concepts like good and bad, right and wrong, virtue, justice, and happiness were the concern of human civilizations through millenniums. Historically the foundations of human ethics are laid by divine revaliations through prophets. Then Greek phlosophers and Muslim scholars contributed to the theory till the begining of the modern times. With the begining of the 20th century industrial society began to transform into information society and risk and uncertainty became prior and diagnostic feature of human behaviour. And now with these changes new structure of society is multi-dimensional, more complicated and uncertain. While this theory became accepted and is using in economics, criticism voices started to become louder. Critics to economics‘ deductive, abstractive and pure rationalist method focused especially on uncertainty and risk that propelled from that analysis. Among them only one was seem as an strong alternative to expected utility theory: Prospect theory which was found by Daniel Kahneman and Amos Tversky in (Kahneman and Tversky 1979). They dealt with utility concept from the cognitive point of view, Daniel Kahneman and Amos Tversky‘s studies were the earliest studies of human decision-making by cognitive psychologists. Development of the theory through modern times may be the subject of another work. References Alberti, A., (1995). The Epicurean Theory of Law and Justice, in A. Laks and M. Schofield (eds) Justice and Generosity, Studies in Hellenistic Social and Political Philosophy, Cambridge: Cambridge University Press. Bailey, C. (1928). The Greek Atomists and Epicurus, Oxford: Clarendon Pres. DeWitt, N.,W., (1954). Epicurus and his Philosophy, Minneapolis: University of Minnesota Press. Epicurus, (1926). Epicurus, the Extant Remains, trans. C. Bailey, Oxford: Clarendon Press. Ibn Haldun, (1977). Mukaddime, Ankara: Onur Yayinlari. Ibn Khaldun, (1967). An Introduction to History: The Muqaddimah, London: Routledge and Kegan Paul. Kahneman, D., Tversky A., (1979). Prospect theory: An analysis of decision under risk. Econometrica, 47(2), 263-291. Quinton, A., (1973). Utilitarian Ethics, London: Open Court Rosen, F. (2003). Classical Utilitarianism from Hume to Mill, New York: Routledge Scarre, G. (1994). Epicurus as Forerunner of Utilitarianism, Utilitas, 6:219-31. 485 � Dublin Core The Dublin Core metadata element set is common to all Omeka records, including items, files, and collections. For more information see, http://dublincore.org/documents/dces/. Extent The size or duration of the resource. 252 Title A name given to the resource Comparison of Islamic, Traditional and Alternative Utility Theories Author Author DEMİRSOY, Sümeyye CAN, Mehmet Abstract A summary of the resource. Decision making under uncertainty is always trying to be explained by utility theory. And utility theory‘s roots rely on moral philosophy. Moral philosophy is concerning concepts about good and bad, right and wrong, virtue, justice, etc. It can be say that utilitarianism, which is a field of moral philosophy, is more directly about utility theory. Throughout the human history, from Prophet Abraham to Greek philosophers; Socrates, Aristotle, Epicurus, to Islam scholars al Kindi, al-Farabi, al-Razi, Ibn-i Sina, Ibni Rushd, Ibn-i Haldun, all discussed about ethics and utility concept. Date A point or period of time associated with an event in the lifecycle of the resource 2010-06 Keywords Keywords. Conference or Workshop Item PeerReviewed HB Economic Theory https://omeka.ibu.edu.ba/files/original/f58b35e8a7210776e674d7a7ed5166d5.pdf a89995abd6839c8e0f3623e4e403d800 PDF Text Text 3rd International Symposium on Sustainable Development, May 31 - June 01 2012, Sarajevo http://www.wider.unu.edu/publications/working-papers/discussionpapers/2008/en_GB/dp2008-01/_files/78805634425684379/default/dp2008-01.pdf Narula, R. & Marin, A. (2005). Exploring the relationship between direct and indirect spillovers from FDI in Argentina, Research Memoranda 024, Maastricht : MERIT, Maastricht Economic Research Institute on Innovation and Technology, http://ideas.repec.org/p/dgr/umamer/2005024.html Parto, S. (2008).Innovation and Economic Activity: An Institutional Analysis of the Role of Clusters in Industrializing Economies,Journal of Economic Issues, Available at http://www.accessmylibrary.com/coms2/summary_0286-36151980_ITM. Porter, M. E. (1990), The Competitive Advantages of Nations, Harvard Business Review, March-April, No:2 Porter, M. (2000). Location, Competition and Economic Development: Local Clusters in a Global Economy, Economic Development Quarterley, 14 (1), 15-34 Raco, Mike (1999). Competition, Collaboration and the New Industrial Districts: Examining the Institutional Turn in Local Economic Development, Urban Studies, 36 (5-6): 951-968. Comparison of linear regression and neural network models forecasting tourist arrivals to Turkey Selcuk Cankurt, Abdulhamit Subasi International Burch University, Faculty of Engineering and Information Technologies, Francuske Revolucije bb. Ilidza, Sarajevo, 71000, Bosnia and Herzegovina. E-mail:asubasi@ibu.edu.ba Abstract This paper develops statistical and machine learning methods for estimating tourist arrivals which is one of the donnée for planning the sustainable tourism development. Tourism is arguably one of the world's largest and fastest growing industries. Sustainable tourism 304 �3rd International Symposium on Sustainable Development, May 31 - June 01 2012, Sarajevo development is one of the most promising generators of the sustainable economic development. Realistic tourism projections based on accurate tourism forecasting contribute much for the sustainable tourism development. The challenge of the planning and developing sustainable tourism is to see as the complex paradigm but one of the starting points is the accurate forecasting tourist arrivals. In this study, linear regression and neural network multilayer perceptron (MLP) implementations are considered to make multivariate tourism forecasting for Turkey. Comparison of forecasting performances in terms of correlation coefficient (R), relative absolute error (RAE) and root relative squared error (RRSE) measurements shows that MLP model for regression gives a better performance. Keywords: Tourism forecasting; Tourism demand modelling; Time series; Linear regression; Neural networks; Multilayer perceptron; Multivariate tourism forecasting. 1.INTRODUCTION Tourism demand forecasts are of great economic value both for the public and private sector. Tourism products, such as unfilled airline seats, unoccupied hotel rooms, and unused facilities, cannot be stocked because of their perishable nature (Archer, 1987). Therefore, accurately forecasting tourism demand has great importance to the sectors concerned with tourism, in order to accurate and efficient plans (Petropoulos, Nikolopoulos, & V., 2005; Pai & Hong, 2005). According to the World Travel & Tourism Council (WTTC), travel and tourism is the biggest industry in the world. Since 1992 tourism sector is the largest industry and has the largest employer in the world (Aslan, Alper, Kaplan, Muhittin, Kula, & Ferit, 2008). Turkey's economy grew an average of 6.0% per year in last decade. Currently Turkey is in 16th place on the list of the largest economies of the world and the fastest growing economy among members of the Organization for Economic Cooperation and Development (OECD). The new goals of Turkish tourism were to establish an efficient tourism sector with high international competitiveness while preserving and enhancing of the country’s natural and historical environment and cultural heritage in a sustainable manner (Ministry of Culture, 2007). The statistical methods such as linear regression are suitable for data having seasonal or trend patterns, while artificial neural techniques are also efficient for data which are influenced by the special case, like promotion or extreme crisis (Efendigil, Önüt, & Kahraman, 2009). One major application area of ANNs is forecasting (Gooijer & J., 2006); see (Zhang, Patuwo, & Hu, 1998) and (Hippert, Pedreira, & Souza, 2001). Generally the ANNs are increasingly used to forecast demands for tourism (Law & Au, 1999; Law R. , 2000). (Pattie & Snyder, 1996) used a back-propagation neural network model with two hidden layers to forecast 305 �3rd International Symposium on Sustainable Development, May 31 - June 01 2012, Sarajevo monthly overnight stays in US national park systems. (Law & Au, 1999) presented a feedforward neural network with six input and one output nodes to forecast arrivals in Hong Kong. For more application area of ANN, see (Al-Saba & El-Amin, 1999), (Beccali, Cellura, Lo Brano, & Marvuglia, 2004), (Hobbs, Helman, Jitprapaikulsarn, Konda, & Maratukulam, 1998), (Sozen, Arcaklioglu, & Ozkaymak, 2005), (Sabuncuoglu, 1998), (Vellido, Lisboa, & Vaughan, 1999), (Wong, Lai, & Lam, 2000), (Ayata, Cam, & Yıldız, 2007), (Efendigil, Önüt, & Kahraman, 2009). According to the brief review of literature especially related to tourism demands approaches, this study attempts to develop a multivariate linear regression model and a general regression neural network model for forecasting the number of the tourists coming to Turkey. 2.THEORETICAL BACKGROUND 2.1.Linear regression Multiple linear regression (MLR) attempts to model the linear relationship called the regression function between a dependent variable and more than one independent variables as different from simple linear models with one independent variable. The dependent variable is sometimes also called the predictand, and the independent variables is called the predictors. The model for multiple linear regression, given n observations, is for i = 1,2, ... n. value of predictor, coefficient on the the intercept, also known as the bias in machine learning, predictor, total number of predictors, predictand, error. 2.2 MLP Approach Artificial neural networks (ANNs) (also usually preferred Neural Networks NNs) are computing structures inspired from the biological neural networks. A neural network is made of the interconnected processing units (usually called neurons). They have the ability of learning by adjusting the strength of the interconnections which can be achieved by altering the values called weights through the input data (Haykin S. , 1999). Neuron sums the weighted inputs and conveys the net input through an activation function in order to normalize and produce a result (Jones, 2008). 306 �3rd International Symposium on Sustainable Development, May 31 - June 01 2012, Sarajevo The multilayer network architecture consists of an input layer, two or more hidden layers, and one output layer. Activation function is used for both the hidden and output nodes. While the sigmoid function can be used to squash the output of the neuron to in the hidden layer in order to introduce the non-linearity to NN, linear activation function must use in output layer to predict the numerical values in the regression problems. MLP is trained with supervised learning include the Perceptron learning algorithm, Least-Mean-Squares learning, and Backpropagation. Backpropagation is one of the most popular approximation approaches for training the multilayer feedforward neural networks based on the Widrow–Hoff training rule (Bishop, 1995; Haykin S. , 1999; Aslanargun, Mammadov, Yazici, & Yolacan, 2007). 3.EXPERIMENTAL RESULTS A total of 31 models were obtained on the basis of two regression models and their corresponding parameter selection which are three of them belong to linear regression models and remaining 28 ones belong to MLP models. Those models were evaluated with the validation data through three forecasting accuracy measures: correlation coefficient (R), relative absolute error (RAE), root relative squared error (RRSE). Three linear regression models were examined on the basis of attribute selection parameter: none, M5 and greedy methods. It has been shown that the linear regression model with greedy attribute selection parameter has the best accuracy when you compare with the other linear regression models but also the worst when you compare with MLP regression models. According to result of our linear regression model: 25 attributes don’t affect the results — WEKA builds the regression function by considering the attributes which only statistically contribute to the accuracy of the model (measured in ). It will not consider the attributes that don't contribute the regression equation. So this regression model is telling us that whole sale price of Turkey, consumer prize index of Canada, Denmark, Spain, Russia, number of German, France, Syrian, Poland, Romanian, Norwegian, Switzerlandian visitors, Exchange rate of Russia, Canada, Switzerland don’t affect the arrivals to Turkey. Estimated positive values (coefficients) tell us as value of those attributes increase number of the total visitors. Estimated negative values (coefficients) reduce the result — linear regression model is telling us that the bigger negative value is, the lower the total coming tourist. This can be seen by the negative coefficient in front of the variables. Table 1 Overall performance of linear regression and MLP methods Model Correlation coefficient Relative Root relative absolute error squared error Linear Regression 0.978 18.73% 307 20.70% �3rd International Symposium on Sustainable Development, May 31 - June 01 2012, Sarajevo MLP Regression 0.9874 14.17% 15.86% Figure 1 Comparison of MLP and linear regression methods Among the MLP regression models presented, the best forecasting accuracy was the MLP model composed of three hidden layers with the neuron numbers of 30, 15 and 10 (abbreviated as 30-15-10). In this model the learning rate 0.03, momentum 0.8, epoch 500 values are used and backpropagation training algorithm, sigmoid activation function for hidden nodes and unthresholded linear activation function for output node are employed. It showed R 0.9874, RAE 14.17% and RRSE 15.86% accuracy results. Results obtained from the experiments in this study, support the discussions in the literature reviews topic of this paper. As seen in the table (1) apparently, machine learning MLP regression model have better performance than statistical linear regression model. 4.CONCLUSIONS This study presents a multivariate time-series forecasting to predict the tourism demand to Turkey by employing linear regression and multilayer perceptron methods. The real data sets respect to Turkey and its top ranked 24 tourism clients of the countries are used to compare 308 �3rd International Symposium on Sustainable Development, May 31 - June 01 2012, Sarajevo the performance of the those methods and to find out the achievement of them on forecasting tourism demand to Turkey. Comparison of the experimental results among linear regression and MLP demonstrated that the MLP method had better forecasting accuracy. Experimental results showed that the MLP model can produce lower prediction error and higher prediction accuracy and outperformed the linear regression model. According to the experiments, it can be concluded that the tuned MLP method with the multivariate time series has enough satisfactory to forecast the tourism demand to Turkey. In this study, linear regression model with greedy attributes selection method and MLP (30:15:10) models have shown better performance when compared with other corresponding models in forecasting the number of monthly tourist arrivals to Turkey owing to the RAE and the RRSE measures. Unfortunately, there is no certain or systematic method to select the appropriate model. Our studies showed that among the methods mentioned above MLP regression has better performance but still we need numerous experiments to evaluate and find out the most suitable MLP regression model which can be employed on the multivariate time series forecasting. REFERENCES Al-Saba, T., & El-Amin, I. (1999). Artificial neural networks as applied to long-term demand forecasting. Artificial Intelligence in Engineering. Archer, B. (1987). Demand Forecasting and Estimation. Travel, tourism, and hospitality research. A handbook for managers and researchers pp. 77-85 . Aslan, Alper, Kaplan, Muhittin, Kula, & Ferit. (2008). Approach, International Tourism Demand for Turkey: A Dynamic Panel Data. Munich Personal RePEc Archive MPRA Paper No. 10601. Aslanargun, A., Mammadov, M., Yazici, B., & Yolacan, S. (2007). Comparison of ARIMA, neural networks and hybrid models in time series: tourist arrival forecasting. Journal of Statistical Computation and Simulation Vol. 77, No. 1, January , 29–53. Ayata, T., Cam, E., & Yıldız, O. (2007). Adaptive neuro-fuzzy inference systems (ANFIS) application to investigate potential use of natural ventilation in new building designs in Turkey. Energy Conversion and Management, 48, 1472–1479. Beccali, M., Cellura, M., Lo Brano, V., & Marvuglia, A. (2004). Forecasting daily urban electric load profiles using artificial neural networks. Energy Conversion and Management. Bishop, C. (1995). Neural Networks for Pattern Recognition. 309 �3rd International Symposium on Sustainable Development, May 31 - June 01 2012, Sarajevo Choy, K. L., Lee, W. B., & Lo, V. (2003). Design of an intelligent supplier relationship management system: A hybrid case based neural network approach. Expert Systems with Applications, 24, 225–237. Efendigil, T., Önüt, S., & Kahraman, C. (2009). A decision support system for demand forecasting with artificial neural networks and neuro-fuzzy models: A comparative analysis. Expert Systems with Applications 36 6697–6707. Gooijer, J. G., & J., R. (2006). 25 years of time series forecasting. Hyndman International Journal of Forecasting 22 (2006) 443– 473. H.Witten, I., & Frank, E. (2005). Data Mining: Practical Machine Learning Tools and Techniques (Second b.). New York: Elsevier. Haykin, S. (1999). Neural Networks: a comprehensive foundation. (Second Edition b.). Prentice Hall. Hippert, H. S., Pedreira, C. E., & Souza, R. C. (2001). Neural networks for short-term load forecasting: A review and evaluation. IEEE Transactions on Power Systems, 16, 44–55. Hobbs, B. F., Helman, U., Jitprapaikulsarn, S., Konda, S., & Maratukulam, D. (1998). Artificial neural networks for short-term energy forecasting: Accuracy and economic value. Neurocomputing, 23, 71–84. Jones, M. T. (2008). Artificial Intelligence: A Systems Approach. INFINITY SCIENCE PRESS LLC. Law, R. (2000). Back-propagation learning in improving the accuracy of neural networkbased tourism demand forecasting. Law, R., & Au, N. (1999). A Neural Network Model to Forecast Japanese Demand for Travel to Hong Kong. Tourism Management. Mark Hall, E. F., Pfahringer, B., Reutemann, P., & Witten, I. H. (2009). The WEKA Data Mining Software: An Update; SIGKDD Explorations. 11(1). Ministry of Culture, T. (2007). Tourism strategy of Turkey – 2023. Ankara: Republic of Turkey of Ministry of Culture & Tourism. Pai, P.-F., & Hong, W.-C. (2005). An Improved Neural Network Model in Forecasting Arrivals. Annals of Tourism Research, Vol. 32, No. 4, pp. 1138–1141, Elsevier. Palmer, A., Montano, J. J., & Sese, A. (2006). Designing an artificial neural network for forecasting tourism time series. Tourism Management 27 781–790. Pattie, D., & Snyder, J. (1996). Using a Neural Network to Forecast Visitor Behavior. Annals of Tourism Research. Petropoulos, C., Nikolopoulos, K., & V., A. P. (2005). A technical analysis approach to tourism demand forecasting. Applied Economics Letters 12, 327–333. 310 �3rd International Symposium on Sustainable Development, May 31 - June 01 2012, Sarajevo Reinsel, G. C. (2003). Elements of multivariate time series analysis. Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning internal representations by error propagation. Parallel distributed processing (pp. 318–362). Cambridge, MA: MIT Press. Sabuncuoglu, I. (1998). Scheduling with neural networks: A review of the literature and new research directions. Production Planning and Control, 9(1), 2–12. Sozen, A., Arcaklioglu, E., & Ozkaymak, M. (2005). Turkey’s net energy consumption. Applied Energy, 81(2), 209–221. . Vellido, A., Lisboa, P. J., & Vaughan, J. (1999). Neural networks in business: A survey of applications (1992–1998). Expert Systems with Applications, 17, 51–70. Witt, S. F., & Witt, C. A. (1995, 3). Forecasting tourism demand: a review of empirical research, International Journal of Forecasting. Wong, B. K., Lai, S. V., & Lam, J. (2000). A bibliography of neural network business applications research: 1994–1998. Computers & Operations Research 27,1045–1076. Zhang, G., Patuwo, B. E., & Hu, M. Y. (1998). Forecasting with artificial networks: The state of the art. International Journal of Forecasting, 14, 35– 62. Informatisation of the Judiciary in BiH: Success Factors Nedim Fisekovic, Meliha Handzic International Burch University, Bosnia and Herzegovina Abstract Informatisation of the judicial system covers all aspects of information and communication technology (ICT), including: equipping the courts with modern information technology equipment (desktop computers, servers, and printers), setting up a local area network (LAN) and wide area network (WAN), establishing a system of electronic mail for the judicial institutions, developing a system for case management system (CMS and TCMS), developing and establishing judicial web sites (web portal), computer education of all employees in the courts and prosecution offices, internet access for all users in the judiciary and many more. 311 � Dublin Core The Dublin Core metadata element set is common to all Omeka records, including items, files, and collections. For more information see, http://dublincore.org/documents/dces/. Extent The size or duration of the resource. 1179 Title A name given to the resource Comparison of linear regression and neural network models forecasting tourist arrivals to Turkey Author Author Selcuk , Cankurt Subasi, Abdulhamit Abstract A summary of the resource. This paper develops statistical and machine learning methods for estimating tourist arrivals which is one of the donnée for planning the sustainable tourism development. Tourism is arguably one of the world's largest and fastest growing industries. Sustainable tourism development is one of the most promising generators of the sustainable economic development. Realistic tourism projections based on accurate tourism forecasting contribute much for the sustainable tourism development. The challenge of the planning and developing sustainable tourism is to see as the complex paradigm but one of the starting points is the accurate forecasting tourist arrivals. In this study, linear regression and neural network multilayer perceptron (MLP) implementations are considered to make multivariate tourism forecasting for Turkey. Comparison of forecasting performances in terms of correlation coefficient (R), relative absolute error (RAE) and root relative squared error (RRSE) measurements shows that MLP model for regression gives a better performance. Keywords: Tourism forecasting; Tourism demand modelling; Time series; Linear regression; Neural networks; Multilayer perceptron; Multivariate tourism forecasting. Date A point or period of time associated with an event in the lifecycle of the resource 2012-05-31 Keywords Keywords. Conference or Workshop Item PeerReviewed H Social Sciences (General) https://omeka.ibu.edu.ba/files/original/8e11fd53365bd7b6e585e97d9f3f01ee.pdf ceec5c894426c2e35cd438d66a9c1e99 PDF Text Text Comparison of Machine Learning Algorithms in Recognation of Regulatory Region of DNA Günay Karlı, Şenol Doğan, Faculty of Engineering and Information Technology, International Burch Universiy, Sarajevo, BIH E –mails: gkarli@ibu.edu.ba – gkarli@yahoo.com, sdogan@ibu.edu.ba Keywords: algorithms. Data mining, machine learning, supervised learning, classification, rule-based Abstract Data mining has become an important and active area of research because of theoretical challenges and practical applications associated with the problem of discovering interesting and previously unknown knowledge from very large real world database. These databases contain potential gold mine of valuable information, but it is beyond human ability to analyze massive amount of data and elicit meaningful patterns by using conventional techniques. In this study, DNA sequence was analyzed to locate promoter which is a regulatory region of DNA located upstream of a gene, providing a control point for regulated gene transcription. In this study, some supervised learning algorithms such as artificial neural network (ANN), RULES-3 and newly developed keREM-IREM rule induction algorithms were used to analyse to DNA sequence. In the experiments different option of keREM, RULES-3 and ANN were used, and according to the empirical comparisons, the algorithms appeared to be comparable to well-known algorithms in terms of the accuracy of the extracted rule in classifying unseen data. 516 �1.INTRODUCTION Data mining is the process of finding hidden patterns from data. It has wide range of applications such as, predicting stock prices, identifying suspected terrorists and scientific discovery like analysis of DNA microarray (Hanuman et al., 2009) Researchers can now routinely investigate the biological molecular state of a cell measuring the simultaneous expression of tens of thousands of genes using DNA microarrays (Shelke and Deshmukh, 2007). Datamining can be used in the classification of proteins by basing on its primary structures (sequences) is presented. It contains four steps which include textmining, feature selection, datamining and classification. The sequences of protein are collected in a file (Mhamdi and Elloumi, 2004). Artificial neural networks are among the newest signal-processing technologies in the engineer's toolbox. The field is highly interdisciplinary. In engineering, neural networks serve two important functions: as pattern classifiers and as nonlinear adaptive filters. Well-known ANN algorithms are based on the notion of perceptron (Rosenblatt, 1962). Perceptrons can only classify linearly separable sets of instances. If a straight line or plane can be drawn to separate the input instances into their correct categories, input instances are linearly separable and the perceptron will find the solution. If the instances are not linearly separable learning will never reach a point where all instances are classified properly. Multilayered Perceptrons (Artificial Neural Networks) have been created to try to solve this problem (Rumelhart et al., 1986). Properly determining the size of the hidden layer is a problem, because an underestimate of the number of neurons can lead to poor approximation and generalization capabilities, while excessive nodes can result in overfitting and eventually make the search for the global optimum more difficult. An excellent argument regarding this topic can be found in (Camargo and Yoneyama, 2001). Kon & Plaskota also studied the minimum amount of neurons and the number of instances necessary to program a given task into feedforward neural networks (Kon and Plaskota, 2000). There are several algorithms with which a network can be trained (Neocleous and Schizas, 2002). However, the most wellknown and widely used learning algorithm to estimate the values of the weights is the Back Propagation (BP) algorithm. Feed-forward neural networks are usually trained by the original back propagation algorithm or by some variant. Their greatest problem is that they are too slow for most applications. One of the approaches to speed up the training rate is to estimate optimal initial weights (Yam and Chow, 2001). Another method for training multilayered feedforward ANNs is Weight-elimination algorithm that automatically derives the appropriate topology and therefore avoids also the problems with overfitting (Weigend et al., 1991). Genetic algorithms have been used to train the weights of neural networks (Siddique and Tokhi, 2001) and to find the architecture of neural networks (Yen and Lu, 2000). There are also Bayesian methods in existence which attempt to train neural networks. Vivarelli & Williams compare two Bayesian methods for training neural networks (Vivarelli and Williams, 2001). 517 �In recent years, there has been a growing amount of research on inductive learning. In its broadest sense, induction (or inductive inference, supervised learning) is a method of moving from the particular to the general - from specific examples to general rules (Quinlan, 1986). Induction can be considered the process of generalizing a procedural description from presented or observed examples. The purpose of inductive learning is to perform a synthesis of new knowledge, and this is independent of the form given to the input information. RIPPER is a well-known rule-based supervised learning algorithm (Cohen, 1995). It forms rules through a process of repeated growing and pruning. Other fundamental learning classifiers based on decision rules include the AQ family (Michalski and Chilausky, 1980) and CN2 (Clark and Niblett, 1989). Bonarini gave an overview of fuzzy rule-based classifiers (Bonarini, 2000). Fuzzy logic tries to improve classification and decision support systems by allowing the use of overlapping class definitions. Furnkranz (2001) investigated the use of round robin binarization (or pairwise classification) as a technique for handling multi-class problems with separate and conquer rule learning algorithms. The PART (Frank and Witten, 1998) algorithm infers rules by repeatedly generating partial decision trees, thus combining the two major paradigms for rule generation − creating rules from decision trees and the separate-and-conquer rule learning technique. RULES family algorithms (Aksoy, 1993) obtain the IF-THEN rules from a given set of examples. REX-1 (Akgöbek et al., 2006) uses the entropy value to give a greater priority to the attributes with higher importance and obtain more general rules. Segments of genome coding for messenger ribonucleic acids (mRNAs), transfer ribonucleic acids (tRNAs), ribosomal ribonucleic acids (rRNAs) are called genes. Among these mRNAs determine the sequence of amino acids in proteins. The mechanism is simple for the prokaryotic cell where all the genes are converted into the corresponding mRNA (messenger ribonucleic acid) and then into proteins. Genome analysis (Gene finding) typically refers to the area of computational biology that is concerned with algorithmically identifying stretches of sequence, usually genomic DNA, that are biologically functional. This especially includes protein-coding genes, but may also include other functional elements such as RNA genes and regulatory regions. Gene finding is one of the first and most important steps in understanding the genome of a species. Computational Gene prediction is relatively simple for the prokaryotes where all the genes are converted into the corresponding mRNA and then into proteins. The process is more complex for eukaryotic cells where the coding DNA sequence is interrupted by random sequences called introns. Some of the questions which biologists want to answer today are (Jayaram and Bhushan, 2000).: Given a DNA sequence, what part of it codes for a protein and what part of it is junk DNA. 518 �Classify the junk DNA as intron, untranslated region, transposes, dead genes, regulatory elements etc. Divide a newly sequenced genome into the genes (coding) and the non-coding regions. In this study, short sequence of DNA is used as an example set to train the keREM, ANN and RULES-3. The features of the DNA sequence are the nucleotides (a,g,c,t). The learning system is requested to generate a classifier that identifies these sequences whether or not they are in the one of functional DNA regions (coding regions). 2.INTRODUCTION TO GENE A gene is a segment of nucleic acid that contains the information necessary to produce a functional product, usually a protein. Genes consist of a long strand of DNA (RNA in some viruses) that contains a promoter, which controls the activity of a gene, and a coding sequence, which determines what the gene produces. The genes are made up of a coding alphabet of 4 nucleotides made up of 4 bases: Adenine(A), Thymine (T), Guanine (G) and Cytosine (C). The bases Adenine (A) and Guanine (G) are Purines; while Thymine (T) and Cytosine (C) are Pyrimidines. 2.1. Universal Genetic Code There four bases Adenine (A), Thymine (T), Guanine (G) and Cytosine (C) As there are 20 amino acids if we use 2 codons for an amino acid we will short of the representation as 42 = 16. So we use three codons to represent all the 20 amino acids as 43=64. As there are only 20 amino acids and 64 codon representation most of the codon are degenrative. 'ATG' is the start codon and TAG, TGA, TAA are stop codons usually. 2.2. Genome Organisation Genome organization refers to the sequential, not the structural organization of the genome. Besides the coding exons, the non-coding DNA in Eukaryotes may fall in the following classes. Introns: They are DNA sequences inserted between the exons and found in the ORF (Open Reading Frames). They are spliced after the first level of transcription. Most introns are junk inserted within genes. Pseudogenes. 'Dead', non-functional copies of genes present elsewhere in the genome, but no longer of any use. 519 �Retropseudogenes: Like pseudogenes, but have been processed, i.e. lack introns produced by the action of reverse transcriptase (RT) on mRNA, and subsequent incorporation of the cDNA into the genome. Transposons: Jumping genes, which splice themselves in and out of the genome (in DNA form) randomly, by the action of transposase. Retrotransposons: Transcribed into an mRNA, which encodes an RT enzyme, which then copies the mRNA back to DNA and incorporates it into the genome. In fact in humans only 1.5% of the entire genome length corresponds to coding DNA. This 1.5% codes for about 27,000 genes, which in turn code for proteins that are responsible for all the cellular processes 2.3. What are Promoters? A promoter is a regulatory region of DNA located upstream (towards the 5' region) of a gene, providing a control point for regulated gene transcription. 3 . ARTIFICIAL NEURAL NETWORK (ANN) Information processing paradigm that is inspired by the way biological nervous systems, such as the rain, process information. The key element of this paradigm is the novel structure of the information processing system. It is composed of a large number of highly interconnected processing elements (neurones) working in unison to solve specific problems. ANNs, like people, learn by example. An ANN is configured for a specific application, such as pattern recognition or data classification, through a learning process. Learning in biological systems involves adjustments to the synaptic connections that exist between the neurones. This is true of ANNs as well (Domingos, 1995). There are different types of neural networks, which can be distinguished on the basis of their structure and directions of signal flow. Each kind of neural network has its own method of training. Generally, neural networks may be differentiated as follows. feedforward networks (one-layer networks and multi-layer networks ) recurrent networks cellular networks Principles of training multi-layer neural network use backpropagation. 520 �The project describes teaching process of multi-layer neural network employing backpropagation algorithm. To illustrate this process the three layer neural network with two inputs and one output, which is shown in the picture below, is used: Figure 1: Multi-Layer Neural Network Each neuron is composed of two units. First unit adds products of weights coefficients and input signals. The second unit realise nonlinear function, called neuron activation function. Signal e is adder output signal, and y = f(e) is output signal of nonlinear element. Signal y is also output signal of neuron. Figure 2: Teaching Process of Multi-Layer NN To teach the neural network we need training data set. The training data set consists of input signals (x1 and x2) assigned with corresponding target (desired output) z. The network training is an iterative process. In each iteration weights, coefficients of nodes are modified using new data from training data set. Modification is calculated using algorithm described below: Each teaching step starts with forcing both input signals from training set. After this stage we can determine output signals values for each neuron in each network layer. 4.INDUCTIVE LEARNING Machine learning algorithms automatically builds a classifier by learning the characteristics of the categories from a set of classified documents, and then uses the classifier to classify documents into predefined categories. (Khan et al., 2010). In recent years, there has been a growing amount of research on inductive learning. In its broadest sense, induction (or inductive inference) is a method of moving from the particular to the general from specific examples to general rules. Michalski explains inductive learning as: Induction can be considered the process of generalizing a procedural description from presented or observed examples The purpose of inductive learning is to perform a synthesis of new knowledge, and this is independent of the form given to the input information. Inductive learning includes learning from examples and learning from observation and discovery. In order to form a knowledge base using inductive learning, the first task is to collect a set of representative examples of expert decisions. Each example belongs to a known class (for example, + or -) and is described in terms of a number of attributes, (for example "hair" or "eyes"). These examples may be specified by an expert as a good tutorial set, or may come from some neutral source such as an archive. The induction process will 521 �attempt to find a method of classifying an example, again expressed as a function of the attributes that explains the training examples and that may also be used to classify previously unseen cases. (Quinlan, 1993).. The outcome of an induction algorithm is either a decision tree or a set of rules. (Quinlan, 1986).. 4.1. RULES-3 Inductive Learning Algorithm RULES-3 (Aksoy, 1993) is a simple algorithm for extracting a set of classification rules from a collection of examples for objects belonging to one of a number of known classes. An object must be described in terms of a fixed set of attributes, each with its own range of possible values, which could be nominal or numerical. For example, attribute "length" might have nominal values {short, medium, long} or numerical values in the range {-10, 10}. An attribute-value pair constitutes a condition in a rule. If the number of attributes is Na, a rule may contain between one and Na conditions. Only conjunction of conditions is permitted in a rule and therefore the attributes must be all different if the rule comprises more than one condition. This algorithm can be summarized as follows: Step1. Define ranges for the attributes, which have numerical values and assign labels to those ranges. Step2. Set the minimum number of conditions (Ncmin) for each rule. Step3. Take an unclassified example. Step4. Nc=Ncmin-l Step5. If Nc<Na then Nc=Nc+i Step6. Take all values or labels contained in the example. Step7. Form objects which are combinations of Nc values or labels taken from the values or labels obtained in Step6. Step8. If at least one of the objects belongs to a unique class then form rules with those objects; ELSE go to Step5. Step9. Select the rule, which classifies the highest number of examples. Step10. Remove examples classified by the selected rule. Step11. If there are no more unclassified examples, then STOP; ELSE go to Step3. Here Nc is the number of condition(s) for each rule and Na is the number of attributes for each example. 522 �4.2. keREM Inductive Learning Algorithm In this section, an algorithm devised for Inductive Learning, newly developed keREM (Inductive Rule Extraction Method) is introduced, which was developed to obtain the IFTHEN rules from a given set of examples. It discards the pitfalls encountered in the some inductive learning algorithms. It uses the value of gain function, to give a greater priority to the attributes with higher importance and obtain rules that are more general. The algorithm can be summarized as follows: Step1. In a given training set, probability distribution and class distribution rate of the each attribute-value pairs is computed. Step2. Power of classification is computed for each attribute in the data set. Step3. Class-based Gain of the each attribute-value pairs is calculated by using computed probability distributions, class distribution rate and power of classification. Step4. Any value of which probability distributions one for n=1 can be selected as a rule. The attribute-values are converted into rules. The classified examples are marked. Step5. Go to step8. Step6. Beginning from the first unclassified example, combinations with n values are formed by taking the attribute-values whose gain is bigger. Step7. Each combination is applied to all of the examples in the set of examples. From the values composed of n combinations, those matching with only on class are converted into a rule. The classified examples are marked. Step8. If all of the examples in the training set are classified then go to step11. Step9. Perform n=n+1 expression. Step10. If n<N the go to step6 Step11 if there is more than one rule representing the same examples, the most general one is selected. Step12. End. 4.3. IREM Inductive Learning Algorithm Newly developed IREM (Inductive Rule Extraction Method) is introduced In this section, which was developed to obtain the IF-THEN rules from a given set of examples. It uses the class-based entropy value, to give a greater priority to the attributes with higher importance and obtain rules that are more general. The algorithm can be summarized as follows: 523 �Step1. In a given training set, probability distribution of the each attribute-value pairs is computed. Step2. The entropy is computed for each attribute and value. Step3. By using computed probability distributions and entropy, class-based entropy is calculated. Step4. Any value of which class-based entropy equals zero for n=1 can be selected as a rule. The values are converted into rules. The classified examples are marked. Step5. Go to step8. Step6. Beginning from the first unclassified example, combinations with n values are formed by taking the value of the attributes whose class-based entropy is smaller. Step7. Each combination is applied to all of the examples in the set of examples. From the values composed of n combinations, those matching with only on class are converted into a rule. The classified examples are marked. Step8. If all of the examples in the training set are classified then go to step11. Step9. Perform n=n+1 expression. Step10. If n<N the go to step6 Step11 if there is more than one rule representing the same examples, the most general one is selected. Step12. End. 5. EXPERIMENTS Using ANN and RULES-3 systems, experiments were performed for the classification of promoter DNA region. For this purpose promoter data-sets available in the University of California-Irvine’s Repository of Machine Leraning Databases were used (Merz and Murphy, 1996). 5.1. Promoter Recognition Experiments with ANN Using different option of ANN, 30 sets of experiment were performed on the promoter data set of E.coli DNA. 86 (approximately %80 of the data set.) randomly selected instances of the original data set were used as a training data in the experiments (Nayır and Karlı, 2009). In order to built, train and test an ANN to recognize promoter region of the DNA, the above Matlab code was written. By changing “newff()” function’s arguments, some ANN with different number of hidden layer and neurons were designed and tested. Using the code above, an ANN with two hidden layers, the first hidden layer with neuron number of 10 and 524 �the second hidden layer with neuron number of 5 was designed. Activation functions of the input layer and hidden layers were ‘tansig’. And activation function of the output layer was ‘purelin’. Minimum error rate (%95) was gained in this option of the ANN. The following table represents error rate of the ANN with one, two and three hidden layers in our experiments. Figure 3: Performance of the NN at 100 epochs. Figure 4: Performance of the NN at 694 epochs. Table 1: Experiment results of ANN with one hidden layer. 525 �No of set of Layer Exp. 1 2 3 The numb Epoc er of h neuro n Input 10 1.Hidde n 5 Output 1 Input 50 1.Hidde n 25 Output 1 Input 100 1.Hidde n 50 Output 1 Min. error rate 947 35 244 25 1404 20 Table 2: Experiment results of ANN with two hidden layer No of set of Layer Exp. The Min. number Epoch error of rate neuron Input 25 1.Hidden 10 2.Hidden 5 Output 1 1 526 694 5 �2 Input 50 1.Hidden 25 2.Hidden 10 Output 1 Input 100 1.Hidden 50 2.Hidden 25 Output 1 3 851 25 554 10 Table 3: Experiment results of ANN with three hidden layer. No of set of Layer Exp. 1 2 527 The numb Epoc er of h neuro n Input 25 1.Hidde n 10 2.Hidde n 5 3.Hidde n 3 Output 1 Input 50 1.Hidde n 25 Min. error rate 785 45 978 20 �3 2.Hidde n 10 3.Hidde n 5 Output 1 Input 100 1.Hidde n 50 2.Hidde n 25 3.Hidde n 10 Output 1 889 30 5.2. Promoter Recognition Experiments with RULES-3 Using different values of number of condition, three sets of experiments were performed on the promoter data set. 40 randomly selected instances of the original data set were used as a training data in the experiments. When Number of condition was set to 1, rules were produced (Karlı, 2000). Using the extracted rule, 11 instances could not be recognized. So accuracy on test data was 90,1. Using different options, three sets of experiments were performed on the promoter data set. As expected, Rules-3 algorithm produced a rule set that classified all training examples correctly. One important conclusion may be driven from table 4.3 is that while number of condition was decreased, extracted rule number decreased. However, accuracy on test data increased. The highest accuracy on test data was gained when number of condition was equal to 1. Table 4: Results for promoter data set with different values of Nc. 528 �Numb Numbe Numbe Accurac Accurac er of r y y of r of conditi exampl extract on on test es on ed training data(%) rules data(%) 3 40 38 100 61,8 5.3 2 40 25 100 76,4 1 40 18 100 90,1 Promoter Recognition Experiments with keREM The most important features of keREM algorithm is that it can compute class-based gain of each attribute-value in a given training set. In this context, first, probability distribution, class distribution rate and power of classification of each nucleotide forming DNA sequence were computed in terms of promoter and non-promoter classes. In the next step, Class-based Gain was computed for each value in the DNA sequence data set by using computed probability distributions, class distribution rate and power of classification. In this way, rules produced by the algorithm were formed by attribute-value whose information value is maximum. The rule set constructed by the method was applied to DNA test set. And the error rate was satisfactory, %97.17 5.4. Promoter Recognition Experiments with IREM The most important features of IREM algorithm is that it can compute class-based entropy of each attribute-value in a given training set. In this context, first, probability distributions of each nucleotide forming DNA sequence were computed in terms of promoter and nonpromoter classes. In the next step, entropy of training set was found. But, the entropy does not contain class information for the value-attribute pairs. Thus, using the entropy of the training set and the probability distributions of the attribute-value, class-based entropy was computed for each value in the DNA sequence data set. In this way, rules produced by the algorithm were formed by attribute-value whose information value is maximum. The rule set constructed by the method was applied to DNA test set. And the error rate was satisfactory, %98.1 529 �6.CONCLUSION Although from a biochemical view point DNA is a complex molecule, from a computer science view point DNA can be considered a very long string over four alphabets A, C, G T. One of the most important step in analysis of a new DNA sequence is finding out whether or not it contains any genes, and if so, determining exactly where they are. For locating functional region in newly sequenced DNA data keREM, IREM, artificial neural network (ANN) and RULES-3 can be used in such a way that known region in mapped sequences is given as input to the systems. Then, the output classifiers of the keREM, IREM, ANN and RULES-3 are used to locate functional regions of newly sequenced data. In this study, keREM, IREM, ANN and RULES-3 were used to locate promoter region of DNA data. Some portions of DNA serve as protein coding regions. However, some portions serve as regulatory markers for the processes that convert coding regions into protein. One of these regulatory regions is a promoter that occurs before coding regions to signal where the transcription process begins. To recognize promoter region, some sort of experiments were performed by using different options of keREM, IREM, ANN, different numbers of hidden layers and neurons, and RULES-3, In the first sort of experiments, one hidden layer with the number of neuron ranging from 5 to 50 was used in ANN. In these experiments, minimum error rate was 15. In the second sort of experiments, two hidden layers were used, the first hidden layer with the number of neuron ranging from10 to 50 and the second hidden layer with the number of neuron ranging from 5 to 25. And minimum error rate of these experiments was 5. This was the best result gained from the experiments. In the last sort of experiments three hidden layers were used, the first hidden layer with the number of neuron ranging from 10 to 50, the second hidden layer with the number of neuron ranging from 5 to 25 and the last hidden layer with the number of neuron ranging from 3 to 10. And minimum error rate was 20. Using different options of RULES-3, three sets of experiments were performed on the promoter data set. As expected, Rules-3 algorithm produced a rule set that classified all training examples correctly. The highest accuracy on test data was gained when number of condition was equal to 1. It is observed that, rules formed by IREM were more general than that of keREM, RULES-3 and ANN. Only 2 examples could not be recognized out of 106 example test set. As a result, it was determined that the error rate of the IREM was lower than the error rate of the keREM, RULES-3 and ANN for DNA sequence test set. Table 5: The errors of some machine learning algorithms on promoter data set. System Errors Comments REX-1 0/106 Inductive L.A 530 �ILA 0/106 Inductive L.A IREM 2/106 Class-based entropy keREM 3/106 Class-based gain KBANN 4/106 A hybrid ML system ANN 6/106 ANN with hidden layer BP 8/106 Standard backpropagation with one layer RULES3 11/106 Nc=1 O'Neill 12/106 Ad hoc tech. from the bio. lit. NearNeigh 13/106 A nearest neighbours algorithm ID3 19/106 Quinlan's builder ANN 21/106 ANN with hidden l. two decision three As it may be driven from table 5, using the same data set, error rate of accuracy on the test data of the REX and ILA is 0/106, which is the minimum error rate. The second best result belongs to IREM with the error rate 2/106. REFERENCES Quinlan, R. (1986). Induction Decision Tree. Machine Learning, 1, 81-106. Jayaram, P., Bhushan, K. (2000). Bioinformatics For Better Tomorrowç. Indian Institute of Technology, Hauz Khas: New Delhi. Domingos, P. (1995). Rule Induction and Instance Based Learning. IJCAI-95. 531 �Clark, P., Niblett, T. (1989). The CN2 Induction Algorithm. Machine Learning, 3, 261-283. Aksoy, M. S. (1993). New Algorithms for Machine Learning, University of Wales: Cardiff. Merz, C. J., Murphy, P. M. (1996). UCI Repository of Machine Learning Database, Retrieved April 1, 2010, from http://www.ics.uci.edu/~mlearning/MLlearning. Nayır, A., Karlı, G. (2009). Application of Artificial Neural Network (ANN) to DNA Sequence Analysis, In AICT 2009, The 3rd IEEE International Conference on Application of Information and Communication Technologies. Karlı, G., (2000). Application of Rule Induction Algorithms to DNA Sequence Analysis. Fatih University: İstanbul. Akgöbek, Ö., Aydın, Y. S., Aksoy, M. S. (2006). A new algorithm for automatic knowledge acquisition in inductive learning. Knowledge-based Systems, 19, 388-395. Shelke, RR., Deshmukh, V. M. (2007). Computational analysis of DNA microarray data using data mining. Biosciences Biotechnology Research Asia, 4, 321-324. Mhamdi, F., Elloumi, M., Rakotomalala, R. (2004). Textmining feature selection and datamining for proteins classification. In ICTTA 2004, International Conference on Information and Communication Technologies: From Theory to Applications, 457-458. Hanuman, T., Raghava, M., Siva, A., Mrithyunjaya, K , Chandra, V. (2009). Performance Comparative in Classification Algorithms Using Real Datasets. Comput Sci Syst Biol, 2, 97100. Khan, A., Baharudin, B., Lee, L. H, Khan, K, A. (2010). Review of Machine Learning Algorithms for Text-Documents Classification, Journal Of Advances In Information Technology. Rosenblatt, F. (1962). Principles of Neurodynamics. Spartan: New York. Rumelhart, D. E., Hinton, G. E., Williams, R. J. (1986). Learning Internal Representations by Error Propagation. In: Rumelhart D E, McClelland J L et al. (eds.) Parallel Distributed Processing: Explorations in the Microstructure of Cognition. MIT Press, Cambridge, MA, 1, 318-362. Camargo, L. S. & Yoneyama, T. (2001). Specification of Training Sets and the Number of Hidden Neurons for Multilayer Perceptrons. Neural Computation 13: 2673–2680. Kon, M. & Plaskota, L. (2000), Information complexity of neural networks, Neural Networks 13: 365–375. Neocleous, C. & Schizas, C., (2002), Artificial Neural Network Learning: A Comparative Review, LNAI 2308, pp. 300–313, Springer-Verlag Berlin Heidelberg. 532 �Yam, J. & Chow, W. (2001). Feedforward Networks Training Speed Enhancement by Optimal Initialization of the Synaptic Coefficients. IEEE Transactions on Neural Networks, 12, 430-434. Weigend, A. S., Rumelhart, D. E., & Huberman, B. A. (1991). Generalization by weightelimination with application to forecasting. In: R. P. Lippmann, J. Moody, & D. S. Touretzky (eds.), Advances in Neural Information Processing Systems 3, San Mateo, CA: Morgan Kaufmann. Siddique, M. N. H. and Tokhi, M. O. (2001), Training Neural Networks: Backpropagation vs. Genetic Algorithms, IEEE International Joint Conference on Neural Networks, 4, 2673–2678. Yen, G. G. and Lu, H. (2000), Hierarchical genetic algorithm based neural network design, IEEE Symposium on Combinations of Evolutionary Computation and Neural Networks, 168– 175. Vivarelli, F. & Williams, C. (2001). Comparing Bayesian neural network algorithms for classifying segmented outdoor images. Neural Networks, 14, 427-437. W. W. Cohen. (1995). Learning to classify English text with ILP methods. In Luc De Raedt, editor, Advances in inductive logic programming, 124–143. IOS Press, Amsterdam, NL. Michalski, R. S., Chilausky, R. L. (1980). Learning by being told and learning from examples: an experimental comparison of the two methods of knowledge acquisition in the context of developing and expert system for soybean disease diagnosis. Policy Analysis and Information Systems, 4. Bonarini, A. (2000), An Introduction to Learning Fuzzy Classifier Systems. Lecture Notes in Computer Science, 1813, 83-92. Furnkranz, J. (1997). Pruning algorithms for rule learning. Machine Learning, 27, 139-171. Frank, E. & Witten, I. (1998). Generating Accurate Rule Sets Without Global Optimization. In Shavlik, J., (eds), Machine Learning: Proceedings of the Fifteenth International Conference, Morgan Kaufmann Publishers. Pham, D.T., Dimov, S.S. (1997). The RULES-4 incremental inductive learning algorithm, in: R.A. Adey, G. Rzevski, R. Teti (Eds.), Applications of Artificial Intelligence in Engineering XII. Computational Mechanics Publications. 533 � Dublin Core The Dublin Core metadata element set is common to all Omeka records, including items, files, and collections. For more information see, http://dublincore.org/documents/dces/. Extent The size or duration of the resource. 1211 Title A name given to the resource Comparison of Machine Learning Algorithms in Recognation of Regulatory Region of DNA Author Author Gunay, Karli Abstract A summary of the resource. Keywords: Data mining, machine learning, supervised learning, classification, rule-based algorithms. Abstract Data mining has become an important and active area of research because of theoretical challenges and practical applications associated with the problem of discovering interesting and previously unknown knowledge from very large real world database. These databases contain potential gold mine of valuable information, but it is beyond human ability to analyze massive amount of data and elicit meaningful patterns by using conventional techniques. In this study, DNA sequence was analyzed to locate promoter which is a regulatory region of DNA located upstream of a gene, providing a control point for regulated gene transcription. In this study, some supervised learning algorithms such as artificial neural network (ANN), RULES-3 and newly developed keREM-IREM rule induction algorithms were used to analyse to DNA sequence. In the experiments different option of keREM, RULES-3 and ANN were used, and according to the empirical comparisons, the algorithms appeared to be comparable to well-known algorithms in terms of the accuracy of the extracted rule in classifying unseen data. Date A point or period of time associated with an event in the lifecycle of the resource 2012-05-31 Keywords Keywords. Conference or Workshop Item PeerReviewed H Social Sciences (General),T Technology (General) https://omeka.ibu.edu.ba/files/original/49c638afaab4f7884c5f976f85b9cc75.pdf aa2806ee70d36f9ce713776c1ae3f233 PDF Text Text COMPARISON OF MACHINE LEARNING TECHNIQUES IN PHISHING WEBSITE CLASSIFICATION Adnan Hodžić International Burch University Bosnia and Herzegovina adnan.hodzic@ibu.edu.ba Jasmin Kevrić International Burch University Bosnia and Herzegovina jasmin.kevric@ibu.edu.ba Adem Karadag Turkey nuhadem@gmail.com Abstract: Phishing is one among the luring strategies utilized by phishing artist in the aim of abusing the personal details of unsuspected clients. Phishing website is a counterfeit website with similar appearance, but changed destination. The unsuspected client post their information thinking that these websites originate from trusted financial institutions. New antiphishing techniques rise continuously, yet phishers come with new strategy by breaking all the antiphishing mechanisms. Hence there is a need for productive mechanism for the prediction of phishing website. This paper described comparison in classification of phishing websites using different Machine­learning algorithms. Random Forest (RF), C4.5, REP Tree, Decision Stump, Hoeffding Tree, Rotation Forest and MLP were used to determine which method provides the best results in phishing websites classification. All instances are categorized as 1 for “Legitimate”, 0 for “Suspicious” and ­1 for “Phishy”. Results show that RF with REP Tree show the best performance on this dataset for classification of phishing websites. Keywords: Machine Learning, Phishing Websites Introduction Internet is not only significant for individual users but also for online business organizations. These organizations usually offer online trading(Liu & Ye, 2003). Nevertheless, Internet­ users can be prone to different types of web­threats that can make financial damages, identity theft, loss of private information, brand reputation damage and loss of client’s trust in e­commerce and online banking. Therefore, Internet appropriateness for commercial sales becomes doubtful. Phishing websites is a semantic intrusion which targets the user instead of computer. It is a fairly new Internet crime when compared to other forms, such as virus and hacking. The phishing problem is a tough problem due to the fact that it is extremely easy for an attacker to make a replica of a good website, which looks very authentic to users. ICESoS 2016 - Proceedings Book 249 �International Conference on Economic and Social Studies (ICESoS’16) Phishing attacks usually aim to acquire confidential information like usernames, passwords and financial IDs by tricking users. Phishing attacks typically start by sending an email that appears to come from authentic company to victims requesting them to update or validate their information by visiting a link within the email. The idea is that bait is dropped out hoping that a user will take it and bite into it just like the fish. Usually, bait is an instant messaging website or an e­mail, which will take the user to hostile phishing websites(James, 2005). The motivation behind this study is to make a strong and effective technique which uses Data Mining algorithms and mechanisms to detect phishing websites. Associative and classification algorithms can be very helpful in identifying Phishing websites. It can give us answers about what the most important phishing website features and indicators are and how they link to each other. Comparing between various Data Mining classification and association systems and techniques is also a goal of this study since there are only few investigations that compares different data mining methods in predicting phishing websites. Literature Review Numerous methodologies are being implemented at present to classify phishing websites.(Aburrous, Alamgir, Keshav, & Fadi, 2009) suggests a method for intelligent phishing detection using fuzzy data mining. In this study, e­banking phishing website detection degree is achieved based on six attributes: URL & Domain Identity, Security and Encryption, Source Code and Java script, Page Style and Contents, Web Address Bar, and Social Human Factor. Fuzzy logic and data mining algorithms are applied to classify e­banking phishing websites. (Basnet, Ram, Srinivas, & Sung, 2008) adopts machine learning way for identifying phishing attacks. Support vector machine, biased support vector machine and neural network are used for the effective prediction of phishing e­mails. The objective of this study is to classify phishing emails by combining basic features in phishing emails and utilizing several machine learning algorithms for the classification process. (Mohammad, Fadi, & Lee, 2013) suggested an intelligent prototype for predicting phishing attacks based on Artificial Neural Network. Same authors shed light on the key features that classify phishing websites from real ones and evaluate how good rule­based data mining classification methods are in detecting phishing websites and which classification approach is proven to be more reliable (Mohammad, Lee, & Fadi, 2014). Methodology ● Dataset Dataset used for the research is “Phishing Websites Data Set” (“UCI Machine Learning Repository: Phishing Websites Data Set,” 2016). This dataset was gathered mainly from: PhishTank archive, MillerSmiles archive, Google’s searching operators. 250 ICESoS 2016 - Proceedings Book �Regional Economic Development: Entrepreneurship and Innovation The authors shed light on the key features that have been proven to be solid and efficient in predicting phishing websites while proposing some new features, experimentally assigning new rules to some well­known features and updating some other features. The dataset is divided into 3 parts, training set and 2 test sets. The training set has 11055 and test sets have 2456 and 2670 instances. All instances are categorized as 1 for “Legitimate”, 0 for “Suspicious” and 1 ­ for “Phishy”. Dataset phishing criteria is divided into 4 sections (Address Bar based Features, Abnormal Based Features, HTML and JavaScript based Features and Domain based Features) and it has 30 attributes. Table 1: Phishing features Features group Features Factor Indicator Using the IP Address Long URL to Hide the Suspicious Part Using URL Shortening Services “TinyURL” URL’s having “@” Symbol Redirecting using “//” Adding Prefix or Suffix Separated by (­) to the Domain Address Bar based Features Sub Domain and Multi Sub Domains HTTPS (Hyper Text Transfer Protocol with Secure Sockets Layer) Domain Registration Length F avicon Using Non­Standard Port The Existence of “HTTPS” Token in the Domain Part of the URL Request URL URL of Anchor Links in <Meta>, <Script> and <Link> tags Abnormal Based Features Server Form Handler (SFH) Submitting Information to Email Abnormal URL Website Forwarding HTML and JavaScript based Features Status Bar Customization Disabling Right Click Using Pop­up Window IFrame Redirection Age of Domain DNS Record Website Traffic Domain based Features PageRank Google Index Number of Links Pointing to Page Statistical­Reports Based Feature ICESoS 2016 - Proceedings Book 251 �International Conference on Economic and Social Studies (ICESoS’16) ● Algorithms Several different machine learning algorithms were used for experiments. 1. Multilayer Perceptron (MLP) Multilayer Perceptron is the most frequently used neural network classifier. MLP is a neural network and a neural network can be described as an artificial neural network which consists of a huge number of interconnected processing components known as neurons that act as a microprocessor. It is a mathematical model for classification of non­linear data into distinct classes. Multilayer Perceptron is the most popular and frequently used neural network design (Bishop, 1995). The MLP is feed­forward network architecture which involves two layers with one or more than one hidden layers; the layers are named as the input layer, hidden layer, the output layer. 2. Random Forest Random forests are a mixture of tree predictors where each tree depends on the values of an arbitrary vector sampled individually and with the same allocation for all trees in the forest. The generalization error for forests converges a.s. to a limit as the amount of trees in the forest becomes great. The generalization error of a forest of tree classifiers hangs on the strength of the individual trees in the forest and the relationship between them (Breiman, 2001). 3. Decision Trees Decision Tree Classification produces the output as a binary tree like construction called a decision tree. A Decision Tree model includes rules to predict the target variable. This algorithm scales well, even where there are changing numbers of training examples and significant numbers of attributes in big databases. a) J48 J48 algorithm is an implementation of the C4.5 decision tree algorithm. J48 uses the greedy technique to induce decision trees for classification (Chen, Zheng, Lloyd, Jordan, & Brewer, 2004). A decision­tree model is built by examining training data and the model is used to classify hidden data b) Reduced­Error Pruning (REPTree) REPTree is a quick decision tree learner. Constructs a decision/regression tree utilizing data gain/variance and prunes it adopting reduced­error pruning (with backfitting). REPTree only sorts values for numeric features once. Missing values are dealt with by splitting the related instances into pieces (i.e. as in C4.5). c) Decision Stump Decision stump is an algorithm for building and using a decision stump. It is typically used in combination with a boosting algorithm. Decision stump algorithm does regression (mean­squared error) or classification (entropy). Missing is handled as a separate value (“DecisionStump”, 2016). 252 ICESoS 2016 - Proceedings Book �Regional Economic Development: Entrepreneurship and Innovation d) Hoeffding Tree A Hoeffding tree (VFDT) is an incremental, anytime decision tree induction algorithm that can learn from great data streams, supposing that the distribution generating examples does not vary over time. Hoeffding trees uses the fact that a small sample can often be adequate to choose a best splitting attribute. This idea is supported by the Hoeffding bound, which quantifies the number of observations (Hulten, Geoff, Laurie, & Pedro, 2001). 4. Rotation Forest Rotation Forest is an ensemble technique which trains L decision trees separately, using a different set of obtained features for each tree. Rotation Forest (Rodriguez, Kuncheva, & Alonso, 2006) draws upon the Random Forest idea. The base classifiers are also separately built decision trees, but in Rotation Forest every tree is trained on the whole data set in a rotated feature space. While the tree learning algorithm constructs the classification regions using hyperplanes parallel to the feature axes, a small rotation of the axes may guide to a very different tree. ● Feature Ranking Feature ranking was applied through WEKA software using Correlation Attribute Evaluation(“CorrelationAttributeEval,” 2016). It evaluates the value of an attribute by measuring the correlation (Pearson’s) between it and the class. Nominal attributes are measured on a value by value basis by regarding each value as an indicator. A general correlation for a nominal attribute is reached at via a weighted average. We selected all attributes whose weight is above 0.1. Those are: ❏ ❏ ❏ ❏ ❏ ❏ ❏ ❏ ❏ ❏ ❏ ❏ HTTPS URL of Anchor Adding Prefix or Suffix Separated by (­) to the Domain DNS Record Sub Domain and Multi Sub Domains Request URL Domain Registration Length Server Form Handler (SFH) Links in <Meta>, <Script> and <Link> tags Google Index Age of Domain PageRank ICESoS 2016 - Proceedings Book 253 �International Conference on Economic and Social Studies (ICESoS’16) Experiments and Results All experiments were conducted in WEKA tool (“Weka 3 ­Data Mining with Open Source Machine Learning Software in Java,” 2016) which is an open source data mining application created in JAVA at Waikato University. Table 2: Full training set results Classifier Test 1 Test 2 MLP 85.5% 85% Random Forest 85.7% 84.5% C4.5 74.6% 73% REPTree 88.4% 88% Decision Stump 86.1% 87% Hoeffding Tree 87.3% 88.4% Rotation Forest (REP Tree) 89.1% 88.5% Rotation Forest (Hoeffding Tree) 88% 84.6% The results show that Rotation Forest algorithm with REP Tree as a classifier give the best results for both test sets with 89.1% and 88.5% accuracy respectively. Other classifiers were not far behind, except C4.5 with 74.6% and 73% for two test sets. After doing the ranking features with Correlation Attribute Evaluation, we applied the same classifiers. The results are very close to the ones with full training set. Surprisingly, MLP results improved for both test sets to 89% and 86.4%. MLP is also the best classifier for first test set with just 0.1% drop in comparison to Rotation Forest with REP Tree results with the full training set. REP Tree was the best classifier for test set 2 with 87.6% correct classification. The drop in correct classification after feature reduction is applied is 1.17%. 254 ICESoS 2016 - Proceedings Book �Regional Economic Development: Entrepreneurship and Innovation Table 3: Reduced training set results Classifier Test 1 Test 2 MLP 89% 86.4% Random Forest 81.8% 80.2% C4.5 73.9% 73% REPTree 87.1% 87.6% Decision Stump 86.1% 87% Hoeffding Tree 82.1% 83.4% Rotation Forest (REP Tree) 88.9% 87% Rotation Forest (Hoeffding Tree) 87.5% 84% If we compare two result tables, we can see that Rotation Forest with REP Tree as a classifier gives the overall best results with 88.37% correct classification, while MLP outshines all other classifiers when feature reduction is applied. Discussion (Mohammad et al., 2014) conducted the similar feature selection where they selected nine features (Request URL, Age of Domain, HTTPS and SSL, Website Traffic, Long URL, Sub Domain and Multi Sub Domain, Adding prefix or Suffix Separated by (−) to Domain, URL of Anchor and Using the IP Address). If we compare their selected attributes with ours, we can see that we share 6 same features: Request URL, Age of Domain, HTTPS and SSL, Sub Domain and Multi Sub Domain, Adding prefix or Suffix Separated by (−) to Domain and URL of Anchor). Moreover, all of the 30 features fall within 4 different feature groups: Address Bar based Features, Abnormal Based Features, HTML and JavaScript based Features, and Domain based Features. However, none of the 12 selected feature falls within “HTML and JavaScript” based Features. This raises the question whether this group of features is relevant in classification of phishing websites. Conclusion Phishing websites detection has gotten a colossal consideration by greater part of the individuals as it serves to recognize the undesirable data and dangers. Hence, the greater part of the analysts focuses in discovering the best classifier for recognizing phishing websites. This work models the phishing website prediction as a classification task and presents the machine learning approach for predicting whether the given website is legitimate website or phishing. Multilayer perceptron, Decision tree classifiers, and Rotation Forest have been applied for training the prediction model. Training set of 11055 and two test sets of 2456 and 2670 instances with 30 attributes have been ICESoS 2016 - Proceedings Book 255 �International Conference on Economic and Social Studies (ICESoS’16) prepared in order to facilitate training and implementation. From the results it has been found that the Rotation Forest algorithm with REP Tree as a classifier and MLP performs the best on a full training and on reduced set, respectively. When training set was reduced from 30 attributes to 12, the overall results for all classifiers dropped for 1.17%. In the meantime, MLP’s overall results increased from 85.5% to 87.7%. It is hoped that more interesting results will follow on further exploration of data. References • Liu, Jiming, and Yiming Ye. E­commerce Agents: Marketplace Solutions, Security Issues, and Supply and Demand. Berlin: Springer, 2001. Print. • Aburrous, M. R., Alamgir, H., Keshav, D., & Fadi, T. (2009). Modelling Intelligent Phishing Detection System for E­ banking Using Fuzzy Data Mining. In 2009 International Conference on CyberWorlds. http://doi.org/10.1109/cw.2009.43 • Basnet, R., Ram, B., Srinivas, M., & Sung, A. H. (n.d.). Detection of Phishing Attacks: A Machine Learning Approach. In Studies in Fuzziness and Soft Computing (pp. 373–383). • Bishop, C. M. (1995). Neural Networks for Pattern Recognition. Oxford University Press. Breiman, L. (2001). Random Forests. Machine Learning, 45(1), 5–32. • Chen, M., Zheng, A. X., Lloyd, J., Jordan, M. I., & Brewer, E. (n.d.). Failure diagnosis usingdecision trees. In International Conference on Autonomic Computing, 2004. Proceedings. http://doi.org/10.1109/icac.2004.1301345 • CorrelationAttributeEval. (n.d.). Retrieved May 9, 2016, from http://weka. sourceforge.net/doc.dev/weka/attributeSelection/CorrelationAttributeEval. html • DecisionStump. (n.d.). Retrieved May 9, 2016, from http://weka.sourceforge.net/ doc.dev/weka/classifiers/trees/DecisionStump.html • Hulten, G., Geoff, H., Laurie, S., & Pedro, D. (2001). Mining time­ changing data streams. In Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining ­KDD ’01. http://doi. org/10.1145/502512.502529 • James, L. (2005). Phishing Exposed. Syngress. • Liu, J., & Ye, Y. (2003). E­Commerce Agents: Marketplace Solutions, Security Issues, and Supply and Demand. Springer. • Mohammad, R. M., Fadi, T., & Lee, M. (2013). Predicting phishing websites based on self­structuring neural network. Neural Computing & Applications, 25(2), 443– 458. • Mohammad, R. M., Lee, M., & Fadi, T. (2014). Intelligent rule­ based phishing websites classification. IET Information Security, 8(3), 153–160. • Rodriguez, J. J., Kuncheva, L. I., & Alonso, C. J. (2006). Rotation Forest: A New Classifier • Ensemble Method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(10), 1619–1630. • UCI Machine Learning Repository: Phishing Websites Data Set. (n.d.). Retrieved May 9, 2016, from https://archive.ics.uci.edu/ml/datasets/Phishing+Websites • Weka 3 ­Data Mining with Open Source Machine Learning Software in Java. (n.d.). Retrieved: May 9, 2016, from http://www.cs.waikato.ac.nz/ml/weka/ 256 ICESoS 2016 - Proceedings Book � Dublin Core The Dublin Core metadata element set is common to all Omeka records, including items, files, and collections. For more information see, http://dublincore.org/documents/dces/. Extent The size or duration of the resource. 3308 Title A name given to the resource COMPARISON OF MACHINE LEARNING TECHNIQUES IN PHISHING WEBSITE CLASSIFICATION Author Author Hodzic, Adnan Kevric, Jasmin Karadag, Adem Abstract A summary of the resource. Abstract: Phishing is one among the luring strategies utilized by phishing artist in the aim of abusing the personal details of unsuspected clients. Phishing website is a counterfeit website with similar appearance, but changed destination. The unsuspected client post their information thinking that these websites originate from trusted financial institutions. New antiphishing techniques rise continuously, yet phishers come with new strategy by breaking all the antiphishing mechanisms. Hence there is a need for productive mechanism for the prediction of phishing website. This paper described comparison in classification of phishing websites using different Machinelearning algorithms. Random Forest (RF), C4.5, REP Tree, Decision Stump, Hoeffding Tree, Rotation Forest and MLP were used to determine which method provides the best results in phishing websites classification. All instances are categorized as 1 for “Legitimate”, 0 for “Suspicious” and 1 for “Phishy”. Results show that RF with REP Tree show the best performance on this dataset for classification of phishing websites. Date A point or period of time associated with an event in the lifecycle of the resource 2016 Keywords Keywords. Conference or Workshop Item PeerReviewed H Social Sciences (General) Dublin Core The Dublin Core metadata element set is common to all Omeka records, including items, files, and collections. For more information see, http://dublincore.org/documents/dces/. Extent The size or duration of the resource. 3464 Title A name given to the resource COMPARISON OF MEDIEVAL AND MODERN METAPHORICAL CONCEPTS Author Author Štrmelj, Lidija Abstract A summary of the resource. This article aims to study emotion metaphors found in selected Chaucer’s Canterbury Tales and compare them with conventional modern metaphors from current dictionaries and other sources, in order to find out whether medieval emotional metaphorical concepts have survived up to the present-day, and if yes, what changes can be perceived in them. The study is based on the cognitive theory of metaphor, as developed by Lakoff and Johnson in “Metaphors We Live By”. Date A point or period of time associated with an event in the lifecycle of the resource 2014 Keywords Keywords. Conference or Workshop Item PeerReviewed PE English https://omeka.ibu.edu.ba/files/original/8d66acc444227b11b0604c66cff36946.pdf 6841d67bcbb7734abb69327017b5db0a Dublin Core The Dublin Core metadata element set is common to all Omeka records, including items, files, and collections. For more information see, http://dublincore.org/documents/dces/. Title A name given to the resource IT Master's Theses Description An account of the resource IT Master's Thesis collection features master's theses authored by graduate students in the Department of Information Technology. Each thesis reflects a significant research effort, combining theoretical knowledge with practical application to address complex challenges in the IT domain. These works demonstrate students’ advanced understanding of information systems, software engineering, data science, cybersecurity, and emerging technologies. The theses serve as a testament to the students' capability to conduct independent research, propose innovative solutions, and contribute to the advancement of the IT field. Subject The topic of the resource IT Master's Thesis Language A language of the resource English Language Author Author IT Department Master’s Students Dublin Core The Dublin Core metadata element set is common to all Omeka records, including items, files, and collections. For more information see, http://dublincore.org/documents/dces/. Title A name given to the resource Comparison of Performanse and Security Aspects of Database Access via Stored Procedures and APIs Author Author Ramiz Šuvalija Abstract A summary of the resource. Modern applications typically get the information in one of two modalities, namely API as an intermediary layer or stored procedures in the same database. The aim of this study is to contrast these methods, mainly performance-wise, and then securitywise, as well as suitability for maintenance as well as scalability. The project will implement identical stored procedures in the PostgreSQL database, and a API backend in Python. Execution time for a query, resource consumption as well as susceptibility to security flaws will be evaluated. The plan is to perform 10 runs for each comparison so as to ensure the obtained results are as accurate as well as dependable as possible. And one of the aims is to devise practical recommendations as to when to apply a stored procedure, and when the API method, where a boundary (equilibrium) has to be drawn between the logic of intermixing in the same data as well as the logic in the app layer. Today with applications being used in distributed environments on a widespread basis, awareness of them is most important in ensuring smooth and effective development of information systems, particularly in those fields where a lot of information has to be processed, such as e-business, banks etc. Keywords Keywords. Stored Procedures, Api, Comparison, Database Access, Postgresql Api Comparison Database Access master’s thesis Postgresql Stored Procedures