Evolutionary Algorithms for Feature Selection in Data Mining

Abstract

Feature selection plays a pivotal role in data mining and machine learning, aiding in enhancing model performance, reducing overfitting, and improving interpretability. Evolutionary algorithms (EAs)—such as Genetic Algorithms (GA), Particle Swarm Optimization (PSO), Ant Colony Optimization (ACO), and their variants—apply bio-inspired search strategies to identify optimal feature subsets from high-dimensional datasets. These algorithms offer a compelling blend of global search capability and adaptability, particularly suited to navigating large and complex search spaces.

PSO and GA have been especially prominent, with PSO used in nearly half of reported swarm-intelligence-based feature selection cases, underscoring its effectiveness across diverse domains—from intrusion detection to gene expression classification . Hybrid methods combining EAs with filter techniques (e.g., χ² statistics or information gain) or local search heuristics have consistently demonstrated enhanced convergence, higher classification accuracy, and more compact feature sets .

Notably, advanced PSO variants—including binary PSO, adaptive PSO with leadership learning, combinatorial PSO, and two dimensional learning frameworks—have improved performance on specific tasks such as intrusion detection and gene expression data selection, achieving high detection rates with minimal features . These techniques balance exploration and exploitation more effectively, addressing issues like premature convergence and scalability.

This overview synthesizes pre-2020 progress by examining EA designs, hybrid strategies, application contexts, strengths, and limitations. It highlights both the versatility and challenges of evolutionary feature selection— illuminating pathways for future enhancements in scalability, convergence reliability, and integration with domain specific methods.