Task decomposition using geometric relation for min-max modular SVMs

The min-max modular support vector machine (M-3-SVM) was proposed for dealing with large-scale pattern classification problems. M-3-SVM divides training data to several sub-sets, and combine them to a series of independent sub-problems, which can be learned in a parallel way. In this paper, we explore the use of the geometric relation among training data in task decomposition. The experimental results show that the proposed task decomposition method leads to faster training and better generalization accuracy than random task decomposition and traditional SVMs.