Effective Large-scale Sample Reduction Strategy Based on Support Vector Machine

Abstract

Training a support vector machine (SVM) on a large-scale sample set is a challenging problem. This paper proposes a sample reduction strategy to pretreat training samples which is realized by a two step procedure: instance reduction and attribute reduction, and the classification model of the SVM is also offered. The experimental results show that the proposed reduction algorithm can effectively remove the nonsupport vector instances and nonessential attributes of the samples, consequently, the whole sample space is simplified and good results are obtained both in training speed and testing precision.