Performance evaluation and analysis of feature selection algorithms

Abstract

Exorbitant data of huge dimensionality is generated because of wide application of technologies nowadays. Intent of using this data for decision-making is greatly affected because of the curse of dimensionality as selection of all features will lead to over-fitting and ignoring the relevant ones can lead to information loss. Feature selection algorithms help to overcome this problem by identifying the subset of original features by retaining relevant features and by removing the redundant ones. This paper aims to evaluate and analyze some of the most popular feature selection algorithms using different benchmarked datasets K-means Clustering, Relief, Relief-F, Random Forest (RF) algorithms are evaluated and analyzed in the form of combinations of different rankers and classifiers. It is observed empirically that the accuracy of the ranker and classifier varies from dataset to dataset. Novel concept of applying Multivariate co-relation analysis (MCA) for feature selection is made and results show improved performance over legacy based feature selection algorithms.