A Nascent Technique for Cultivated Feature Selection using Evolutionary Computation Algorithms

Abstract

Evolutionary computation algorithms are in recent trends for feature selection due to their efficient results as compared to traditional algorithms. The performance of the evolutionary computational algorithm completely depends upon the parameter (like population, number of generations) setting. Existing parameter setting techniques have been developed for specific algorithms. This paper designs a generic pre-processing technique that removes the redundant as well as irrelevant features from the initial given feature set. This work uses Kendall tau to find the non-redundant feature and the Kruskal Wallis for the relevant features. If Kendall tau is not able to select any non-redundant feature then mutual information gain is used to select significant features followed by the Kruskal Wallis to select relevant features. The designed technique has been analyzed on 4 evolutionary computation algorithms, Modified Binary Particle Swarm Optimization, Non-dominated Sorting Genetic Algorithm, Binary Grey Wolf Optimization, and Binary Whale Optimization Algorithm over 6 datasets by using accuracy, the number of features in selected subset, sensitivity, and specificity. The present technique improves the performance of each algorithm in terms of accuracy, sensitivity, and specificity by 1 % on average with a 10 % reduction in the selected feature subset proves the significance of the proposed technique.
 HIGHLIGHTS
 
 This work removes the redundant features from the initial population using Kendall’s tau and mutual information gain to reduce the initial population for any Evolutionary Computation (EC) algorithm
 The relevant features that are highly correlated with class are added back to the initial population to maintain the optimal initial population
 The performance validation has been done using four EC algorithms to show the significance of the work
 
 GRAPHICAL ABSTRACT