Department of Statistics, The Ohio State University
Statistics and Biostatistics Colloquium Series

Many margin-based classification techniques such as support vector machines (SVM) and psi-learning deliver high performance by directly focusing on estimating the decision boundary, as opposed to estimating the conditional probabilities via regression techniques. As a result, multicategory classification is often treated separately from binary classification; no straightforward generalization is possible. In this talk, I will present a novel multicategory generalization particularly for psi-learning and SVM as a by-product. A statistical learning theory for multicategory psi-learning is developed, as well as its computational tools based on differenced convex (d.c.) programming. We examine the operating characteristics of the proposed methodology via numerical examples, and we show that psi-learning outperforms SVM in generalization. Moreover, psi-learning is more robust against extreme observations that are wrongly classified than SVM.