A hybrid framework for optimal feature subset selection

In the context of optimal subset selection, hybrid feature selection approaches play a significant role that has been the topic of a substantial number of studies because of the growing need for data mining applications. In feature subset selection (FSS) problem; there are two significant shortcomings that need to be addressed: At first, finding a suitable filter method that can be reasonably fast and energetically computed for large volume of data, and second, an efficient wrapper strategy that can discriminate the samples over the entire search space in a considerable amount of time. After a study of the shortcomings of individual feature selection methods (filter or wrapper), this paper investigated a new hybrid feature selection approach with conjunction of filter and wrapper method that can take benefit of both ways for a classification problem. The proposed hybrid uses the filter method as conditional mutual information maximization and wrapper method as genetic algorithm to enhance the overall classification performance and speed up the search process to identify the essential features. The proposed method is known as FWFSS. To get rid of meaningless features and determine the biomarkers, wrapper method as genetic algorithm uses the naive Bayes (NB) classifier as a fitness function. The proposed method is verified on the University of California, Irvine (UCI) repository, and microarray datasets. From experimental study, it is observed that our approach outperforms convenient methods regarding classification accuracy, the number of optimal features reported in the literature.