Training support vector machine through redundant data reduction

Abstract

Support Vector Machine (SVM) training in a large data set involves a huge optimization problem to make SVM impractical even for a moderate data set. In this paper, we propose a method to speed SVM training by removing redundant data in the training data set. The training data set is clustered firstly. Then the Fisher Discriminant Ratio is used to find the boundary between the clustered and scattered data points in one cluster, which is computed based on the distance density of data points to the cluster centroid. The clustered data points that lie around the clustering centroid are far away from the separating hyperplane, which are considered as the redundant data points and removed. While the scattered data points are in the outside layer of cluster, which are near the separating hyperplane and considered as having the potential Support Vectors (SVs) and reserved. Several experimental results show that our approach preserve good classification accuracy while the training time is reduced significantly.