Ensemble of feature selection methods: A hesitant fuzzy sets approach

Display Omitted In this paper, hesitant fuzzy sets are utilized for representing ensemble of ranking algorithms (as a relevancy measure) and ensemble of similarity measures (as a redundancy merit) for feature subset selection.In this paper, the well-known CFS merit has been fuzzified with ensemble of feature ranking algorithms and similarity measures.The proposed MRMR-HFS is recmmended when one deals with high dimensional datasets which suffer from small sample sizes. Moreover, it can be used when speed of feature selection process is matter.The proposed MRMR-HFS method can be used when the search space is extremely large and cannot be searched by meta-heuristic algorithms.Several experimental results as well as non-parametric statistical tests confirm the performance of our MRMR-HFS method in the field of feature selection. Recently, there has been a great attention to develop feature selection methods on the microarray high dimensional datasets. In this paper, an innovative method based on Maximum Relevancy and Minimum Redundancy (MRMR) approach by using Hesitant Fuzzy Sets (HFSs) is proposed to deal with feature subset selection; the method is called MRMR-HFS. MRMR-HFS is a novel filter-based feature selection algorithm that selects features by ensemble of ranking algorithms (as the measure of feature-class relevancy that must be maximized) and similarity measures (as the measure of feature-feature redundancy that must be minimized). The combination of ranking algorithms and similarity measures are done by using the fundamental concepts of information energies of HFSs. The proposed method has been inspired from Correlation based Feature Selection (CFS) within the sequential forward search in order to present a robust feature selection tool to solve high dimensional problems. To evaluate the effectiveness of the MRMR-HFS, several experimental results are carried out on nine well-known microarray high dimensional datasets. The obtained results are compared with those of other similar state-of-the-art algorithms including Correlation-based Feature Selection (CFS), Fast Correlation-based Filter (FCBF), Intract (INT), and Maximum Relevancy Minimum Redundancy (MRMR). The outcomes of comparison carried out via some non-parametric statistical tests confirm that the MRMR-HFS is effective for feature subset selection in high dimensional datasets in terms of accuracy, sensitivity, specificity, G-mean, and number of selected features.