Supervised Classifier

Supervised Machine Learning is an important eld with many immediate applications. As a result, there is an increasing number of public domain tools with a diversity of learning approaches. However, very little work has been done to identify which public domain machine learning tools are \best" and on what kind of data. This research is a comparative study of di erent supervised public domainMachine Learning algorithms. It also includes the use of data analysis to explain the performance. Speci c characteristics of twelve Machine Learning classi ers are analyzed and their performance are compared on twenty nine UCI datasets. Data analysis using visualization and feature selection is also performed to con rm hypothesis about algorithm performance and explore interesting structure in the data. The experimental algorithms are categorized based on classi cation performance and reasons for di erences in classi er performance are discussed. It is shown that the variance of algorithms with respect to classi cation accuracy is due to the nature of experimental datasets and the way various algorithms react to the property/characteristics of the data. The di erences in classi cation performance also due to several other reasons: overtting, pruning techniques, and di culty in specifying optimal learning parameters. ii Acknowledgments There are a large number of people who have made valuable contributions to this project. First and foremost I would like to thank my supervisor Dr. Peter Eklund for his insightful guidance and strong leadership and also for encouraging and helping me to overcome all di culties in task ful llment as well as language barriers in writing this thesis. Thanks also to Richard Cole for his friendship and helping me in many ways. I wish to thank Peter Deer for his interests in the project and for his detailed and valuable comments about discussion writing style on rst drafts of this thesis. I am grateful to G. Klanniscek, M. Soklic and M. Zwitter of the UniversityMedical Center, Ljubljana for the use of the medical data and I. Kononenko for it conversion to a form suitable the induction algorithms. I am also grateful to David Aha (UCI) for the compilation and use of the UCI Repository of Machine Learning Databases. I am also grateful to P. Clark and T. Niblett, P. E. Utgo and C. E. Brodley, S. K. Murthy, T. Kohonel, T. W. Rauber, the programming team at the Institute for Parallel and Distributed High Performance System (the University of Stuttgart), the programming team at the Department of Internal Medicine, Electrical Engineering, and Computer Science (the University of Nevada) for creating wonderful software packages and granting free access to Internet users for academic use. As an important part of this acknowledgment, I would like to thank the Australian Government for granting me the postgraduate scholarship through the Australian Agency for International Development (AusAID), which made this study possible. I would like to thank all sta members of the Department of Computer Science, The University of Adelaide for providing me convenient working platform and for all the comprehensive knowledge about Computer Science I obtained. Thank also to all Honors/Masters students, to Andrew Burrow and Simon Pollitt, who o ered me friendship and helped me to overcome di culties in study and real life. iii