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

Area of Science:

• Computer Science
• Machine Learning
• Bioinformatics

Background:

• Feature Selection (FS) is crucial for optimizing classification efficiency and reducing data dimensionality.
• Harris Hawk Optimization (HHO) is a metaheuristic algorithm effective for continuous optimization tasks.
• Existing HHO requires adaptation for binary optimization problems like FS.

Purpose of the Study:

• To develop a novel Multi-Objective Quadratic Binary HHO (MOQBHHO) for feature selection.
• To enhance classification efficiency and reduce feature subsets simultaneously.
• To adapt the HHO algorithm for binary optimization spaces.

Main Methods:

• Implementation of the MOQBHHO technique using K-Nearest Neighbor (KNN) as a wrapper classifier.
• Utilizing crowding distance (CD) as a third criterion to select optimal non-dominated solutions.
• Evaluation on twelve standard medical datasets for performance estimation.

Main Results:

• The proposed MOQBHHO effectively identifies non-dominated feature subsets.
• MOQBHHO demonstrates superior performance compared to MOBHHO-S, MOGA, MOALO, and NSGA-II.
• The method achieves a better trade-off between classification efficiency and feature reduction.

Conclusions:

• MOQBHHO is a highly effective technique for multi-objective feature selection in binary spaces.
• The approach surpasses deep-based FS methods like Auto-encoder and Teacher-Student based FS (TSFS).
• This methodology offers a promising solution for identifying relevant features in medical datasets.