Evolutionary Computation Based Automatic SVM Model Selection

Abstract

SVM performance is very sensitive to the parameter set. In this paper we propose an automatic and effective model selection method. It is based on evolutionary computation algorithms and use recall, precision and error rate estimated by xialpha-estimate as the optimization targets. Optimized by genetic algorithm (GA) or particle swarm optimization (PSO) algorithm, we demonstrate that SVM could automatically select its multiple parameters and optimize them. Experiments results also verify that by optimizing the bounds estimated by xialpha-estimate we could also improve the practical performance.