Simplified Support Vector Machines Via Kernel-Based Clustering

Reduced set method is an important approach to speed up classification process of support vector machine (SVM) by compressing the number of support vectors included in the machine's solution. Existing works find the reduced set vectors based on solving an unconstrained optimization problem with multivariables, which may suffer from numerical instability or get trapped in a local minimum. In this paper, a novel reduced set method relying on kernel-based clustering is presented to simplify SVM solution. This approach is conceptually simpler, involves only linear algebra and overcomes the difficulties existing in former reduced set methods. Experiments on real data sets indicate that the proposed method is effective in simplifying SVM solution while preserving machine's generalization performance.