Feature ranking using Gini index, scatter ratios, and nonlinear SVM RFE

This paper introduces three feature ranking (FR) methods using seven classification benchmarks created by Exhaustive Search (ES) which selected the best feature subsets. Next, three different FR approaches are compared and ranked in respect to the top five best feature subsets for each data set obtained by ES. The results show that Gini index (GI) outperforms scatter ratios for multi-class problems on average. However, for binary classification, scatter ratios shows better performance. Nonlinear support vector machines for recursive feature elimination (NL SVM RFE) was run on three binary benchmarks and it leads to rankings closest to ES on average over GI and scatter ratio methods.