Improving the Run-Time Performance of Multi-class Support Vector Machines

Abstract

In this paper we propose three approaches to speed up the prediction phase of multi-class Support Vector Machines (SVM). For the binary classification the method of partial sum estimation and the method of orthonormalization of the support vector set are introduced. Both methods rely on an already trained SVM and reduce the amount of necessary computations during the classification phase. The predicted result is always the same as when using the standard method. No limitations on the training algorithm, on the kernel function or on the kind of input data are implied. Experiments show that both methods outperform the standard method, though the orthonormalization method delivers significantly better results. For the multi-class classification we have developed the pairwise classification heuristics method, which avoids a lot of unnecessary evaluations of binary classifiers and obtains the predicted class in a shorter time. By combining the orthonormalization method with the pairwise classification heuristics, we show that the multi-class classification can be performed considerably faster compared to the standard method without any loss of accuracy.