On mixtures of linear SVMs for nonlinear classification

In this paper, we propose a new method for training mixtures of linear SVM classifiers for purposes of non-linear data classification. We do this by packaging linear SVMs into a probabilistic formulation and embedding them in the mixture of experts model. The weights of the mixture model are generated by the gating network dependent on the input data. The new mixture of linear SVMs can be then trained efficiently using the EM algorithm. Unlike previous SVM-based mixture of expert models, which use a divide-and-conquer strategy to reduce the burden of training for large scale data sets, the main purpose of our approach is to improve the efficiency for testing. Experimental results show that our proposed model can achieve the efficiency of linear classifiers in the prediction phase while still maintaining the classification performance of nonlinear classifiers.