Identification of Top-Ranked Features Using Consensus Affinity of State-of-the-Art Methods

Abstract

Feature selection is one of the vital preprocessing tasks of biological data mining process. Although plentiful feature selection techniques have been developed for processing small- to large-dimensional dataset in general but most of them are classifier bound rather than selection of a set of generic features. Hence to address this issue, this work is making an effort to reveal generic features with consensus affinity using state-of-the-art methods based on which various classification models produce unbiased and stable results. In particular, here, we have focused on set of significant features from four benchmark microarray datasets selected by state-of-the-art filter methods such as signal-to-noise ratio, significant analysis of microarray, and t-test and then compared with the features selected by preferred evolutionary wrappers such as Particle Swarm Optimization (PASO), Ant Colony Optimization (ACE), and Genetic Algorithms (Gas). The classification accuracies of several classifiers are measured based on the selected features by the above-mentioned filter techniques and evolutionary wrapper techniques.