Optimization issues in data analysis: An analytic center approach to kernel methods.

Abstract

Support vector machines have recently attracted much attention in the machine learning and optimization communities for their remarkable generalization ability. The support vector machine solution corresponds to the center of the largest hypersphere inscribed in the version space. Recently, however, alternative approaches have suggested that the generalization performance can be further enhanced by considering other possible centers of the version space like the center of gravity. However, efficient methods for calculating the center of gravity of a polyhedron are lacking. A center that can be computed efficiently using Newton's method is the analytic center of a convex polytope. We propose an algorithm that finds the hypothesis that corresponds to the analytic center of the version space. We refer to this type of classifier as the analytic center machine (ACM). In this study ACMs have been employed to solve problems in pattern recognition and regression analysis. Preliminary experimental results are presented for which ACMs outperform support vector machines.