Feature ranking for pattern recognition: A comparison of filter methods

This paper presents an approach for comparing various feature ranking (FR) methods. First, six classification benchmarks are created using Exhaustive Search (ES) to select the best feature subsets. The subset selections have been done within double (nested) cross-validation procedures guaranteeing realistic accuracy predictions to unseen examples. Next, seven filter FR approaches are compared and ranked in respect to the top five best feature subsets for each data set. This paper also introduces a method for quantifying and comparing FR results. The results hint that using Gini index or scatter ratios leads to rankings closest to ES on average.