Fast training of support vector machines using top-down kernel clustering

How to deal with the very large database in decision-making applications is a very important issue, which sometimes can be addressed using SVMs. This paper presents a new sample reduction algorithm as a sampling preprocessing for SVM training to improve the scalability. We develop a novel top-down kernel clustering approach which tends to fast produce balanced clusters of similar sizes in the kernel space. Owing to this kernel clustering step, the proposed algorithm proves efficient and effective for reducing training samples for nonlinear SVMs. Experimental results on four UCI real data benchmarks show that, with very short sampling time, the proposed sample reduction algorithm dramatically accelerates SVM training while maintaining high test accuracy.