Modification of the Entire Regularization Path for the Support Vector Machine

Abstract

The support vector machine (SVM) is a well-known approach to solving machine learning problems. The numerical solution of SVM involves solving the dual of a convex optimization problem. Typically, the convex optimization problem contains parameters that have to be selected a priori. Consequently, the numerical determination of the optimal parameters for a given dataset becomes a computationally expensive process. Recently, it is possible to track the solution of the dual problem over the entire range of parameters. However, several numerical issues remain. This project looks into the possibility of improving the reliability and efficiency of the tracking process. We study two possible modifications of an existing entire regularization path algorithm for the support vector machine. We then show the computational results of the two modifications and compare with the original algorithm.