A New Filter Evaluation Function for Feature Subset Selection with Evolutionary Computation

Feature subset selection is an optimization problem to achieve high classification accuracy with low number of features and low computational cost in the area of pattern classi- fication or data mining. There are various approaches to obtain this. Basically a search algorithm is used with a fitness function either based on intrinsic characteristics of the data, known as filter type, or based on classification accuracy of the classifier used, known as the wrapper type, to find out the optimum feature subset. Both the approaches have respective merits and demerits. Though lots of algorithms are developed so far, none of them works equally well for all the data sets, specially for very high dimensional data sets. In this work, a new feature evaluation measure based on the concept borrowed from topic modelling in text processing, has been developed. The proposed measure is used as a fitness function of evolutionary computational search techniques for designing filter type feature subset selection approach. Simulation experiments with various benchmark data sets have been done for assessing the efficiency of the proposed approach in comparison to the popular conventional filter type feature selection algorithms mRMR and CFS. It is found that the proposed approach is better in terms of selecting lesser number of features with comparable classification accuracy. The proposed algorithms work better for higher dimensional features and can be proved as an effective solution of feature selection for very high dimensional data.