A hybrid approach for optimal feature subset selection with evolutionary algorithms

Feature subset selection is very important as a preprocessing step for pattern recognition and data mining problems. The selected feature subset is expected to produce maximum possible classification accuracy with a minimum possible number of features. For optimal feature selection, a suitable evaluation function and an efficient search method are needed. There are two main approaches. In filter approach, the inherent characteristics of the data set is used for feature evaluation while in wrapper approach, the classification accuracy is used as the evaluation function. Both the approaches have relative merits and demerits. In this paper a suitable combination of both filter and wrapper approch is proposed for selection of optimal feature subset with evolutionary algorithm. Correlation based feature selection (CFS) and minimum redundancy and maximum relevance (mRMR) algorithms are used as filter evaluation approach, binary genetic algorithm (BGA) and binary particle swarm optimization (BPSO) are used as evolutionary serach algorithms. The simulation experiments are done with benchmark data sets. The simulation results show that proper hybridization approach is effective in achieving optimal feature subset selection with minimum number of features having high classification accuracy and low computational cost.