This seminar is about a novel type of learning machine called the support-vector machine (SVM). In simple terms, a learning machine is a computer program that automatically improves with experience (i.e., input-output data) in carrying out some task (prediction, classification, making decisions, etc.). The SVM has been reported by several recent empirical studies to match or outperform the other learning and data mining approaches (including artificial neural networks). Theoretically, the SVM algorithm is based on a well-established mathematical framework that provides actual proofs as to why the SVM should (normally) work well. Also, and contrary to what one may think, none of the basic ideas of SVM is new: they have been around for over three decades now (Vapnik and Chervonenkis, 1974).

Yet, despite all the above, the applications of the powerful SVM algorithm are not as widespread and numerous as those of, for instance, artificial neural networks (ANNs). While ANNs have found applications in virtually all disciplines where data analysis is needed (engineering, business, economics, social sciences, etc.), the SVM is still considered as a new technology to be learned, even in some science and engineering areas. Vapnik (2000) has attributed this delay in the acceptance of SVM to the need of 'intellectual determination': After many successful experiments with SVM, researchers became determined in criticism of the classical philosophy of generalization based on the principle of Occam’s razor. This intellectual determination also is a very important part of scientific achievement.

'Intellectual determination' was required for machine learning theorists to accept SVM as a promising learning machine and data mining tool. The same determination will be required for researchers in engineering, applied science, and management science in order for them to accept implementing the SVM approach in their respective application domains and leverage its power. One of the main goals of this seminar is to help achieve such a determination.

The seminar starts by covering the fundamentals of the mathematical theory on which the SVM approach is based, and which is Statistical Learning Theory (SLT). The basic ideas of SVM are then described in the case of binary classification. The discussion starts with the linearly separable problem, and builds up on the same concepts to introduce the nonlinear case. The VC dimension concept and the results of SLT are used to account for the performance of SVM. Application of SVM to regression estimation is discussed briefly. The limitations of SVM and SLT are explained, and the areas where more research is needed are outlined. Toward the end of the seminar, an application of SVM to environmental informatics (nitrogen removal in biological wastewater treatment systems) is presented. While the topic of SVM is generally highly mathematical and a number of equations cannot be avoided, every attempt will be made in this seminar to provide intuitive explanations and interpretations for the various ideas of SVM and SLT.

Aziz Guergachi is an Assistant Professor in the School of Information Technology Management at Ryerson University. He received a Bachelor's Degree in Engineering from École Supérieure d'Ingénieurs de Marseille (France), a Bachelor of Science in Mathematics from Université de Provence (France), and a Ph.D. in Engineering from the University of Ottawa. He has over 10 years of industry experience, as he had held various positions in the areas of software systems analysis, development and implementation, and project management in both the manufacturing and retail sectors. His research interests lie in the area of advanced modelling, analysis and simulation of engineering, business and social systems. He was recently the recipient of the New Opportunities Award of the Canada Foundation for Innovation.