An analytical study of modified multi-objective Harris Hawk Optimizer towards medical data feature selection

Abstract

Dimensionality reduction or Feature Selection (FS) is a multi-target optimization problem with two goals: improving the classification efficiency while simultaneously dropping the characteristics. Harris Hawk Optimization (HHO) is introduced recently to solve different demanding optimization tasks as a metaheuristic tool. The initial HHO is for addressing optimization problems in a continuous environment, but FS is an optimization task in binary space. Therefore, in this article, a Multi-Objective Quadratic Binary HHO (MOQBHHO) technique with K-Nearest Neighbor (KNN) method as wrapper classifier is implemented for extracting the optimal feature subsets. Finally, this study uses the crowding distance (CD) value as a third criterion for picking the best one from the non-dominated solutions. Here, to estimate the performance of the proposed approach, twelve standard medical datasets are considered. The proposed MOQBHHO is compared with MOBHHO-S (using a sigmoid function), multi-objective genetic algorithm (MOGA), multi-objective ant lion optimization (MOALO), and NSGA-II. The experimental findings show that the proposed MOQBHHO finds a set of non-dominated feature subsets effectively in contrast to deep-based FS methods: Auto-encoder and Teacher-Student based FS (TSFS). The presented methodology is found superior in obtaining the best trade-off between two fitness assessment criteria compared to the other existing multi-objective techniques for recognizing relevant features.