Binarization of Metaheuristic Optimization Algorithm: A Comparative Analysis
Metaheuristic algorithms are a kind of optimization algorithms, which are widely used in reducing the number of variables, and choosing the most relevant, and identifiable ones in medium or high dimensional datasets. This topic, known as feature selection is widely researched and applied where binary metaheuristic algorithms are developed using as an evaluator of the quality of the feature subsets machine learning algorithms. The binary metaheuristics could be created by transforming the continuous space of the metaheuristic by using some operators named transfer function, into a binary one, since feature selection is a binary optimization problem. This paper aims to offer a comparative analysis of three distinct binary variants of metaheuristic optimization algorithms utilizing five different groups of transfer functions to assess their predictive efficacy. Four S-shaped, V-shaped, U-shaped, Q-shaped, and two X-shaped types are proposed for the binarization process. Salp Swarm Optimization, Ant Lion Optimization, and Teaching Learning-Based Optimization algorithms are chosen as the foundational algorithms for assessing each transfer function and comparing their impact on key metrics, including accuracy, F-score, sensitivity, average number of selected features, fitness, and time efficiency. The current work utilizes a medical dataset to predict the existence of gallstones. The experiment aims to compare and propose new solutions applied to the feature selection process. The optimal combinations of transfer functions and metaheuristic algorithms are produced using the combination of the Q3 transfer function together with a binary teaching-based learning method and the V2 transfer function combined with binary antlion optimization, achieving an accuracy of approximately 78%. The research emphasizes the necessity of appropriate binarization strategies and provides insights into selecting transfer functions to improve feature selection in medical data analysis.