Maximum-Margin Nearest Prototype Classifiers with the Sum-over-Others Loss Function and a Performance Evaluation

Abstract

This paper investigates margin-maximization models for nearest prototype classifiers. These models are formulated through a minimization problem, which is a weighted sum of the inverted margin and a loss function. It is reduced a difference-of-convex optimization problem, and solved using the convex-concave procedure. In our latest study, to overcome limitations of the previous model, we have revised the model in both of the optimization problem and the training algorithm. In this paper, we propose another revised margin-maximization model by replacing the max-over-others loss function used in the latest study with the sum-over-others loss function. We provide a derivation of the training algorithm of the proposed model. Moreover, we evaluate classification performance of the revised margin-maximization models through a numerical experiment using benchmark data sets of UCI Machine Learning Repository. We compare the performance of our models not only with the previous model but also with baseline methods that are the generalized learning quantization, the class-wise k-means, and the support vector machine.