A Filter Approach to Feature Selection Based on Survival Cauchy-Schwartz Mutual Information

摘要

In many proposed feature selection methods, those based on information theory have become a hot topic. But when calculating Shannon entropy and mutual information, estimating the probability density of dataset inevitably involves man-made interference, estimation error will increase especially when sample size is small. In this paper, we use survival Cauchy-Schwarz mutual information instead of Shannon information, this kind of mutual information does not need to use the probability density function but calculated from sample data directly. Then combined with a multi-objective optimization algorithm (NSGA-II), the SCS-MIFS-ND feature selection algorithm is proposed. Experimental results on UCI datasets show that the method can improve the accuracy of SVM and select a more optimal subset.