MOELIGA: a multi-objective evolutionary approach for feature selection with local improvement

Researchers Leandro Vignolo and Matias Gerard have introduced MOELIGA, a novel multi-objective evolutionary algorithm designed to tackle the critical challenge of feature selection in machine learning. The algorithm employs a genetic approach enhanced with local improvement strategies, where subordinate populations evolve to refine feature subsets. Key innovations include a crowding-based fitness sharing mechanism to maintain diversity, a sigmoid transformation to guide compactness, and a unique geometry-based objective that promotes classifier independence. This combination allows MOELIGA to intelligently navigate the trade-off between minimizing the number of features and maximizing classification accuracy, a fundamental problem in high-dimensional data analysis.

In comprehensive testing, MOELIGA was evaluated on 14 diverse datasets and benchmarked against 11 state-of-the-art feature selection methods. The experimental results demonstrate that MOELIGA consistently identifies significantly smaller subsets of features while achieving classification performance that is either superior or comparable to existing approaches. The paper's findings, detailed across 49 pages with 9 figures and 4 tables, suggest that MOELIGA offers a robust and adaptable solution for complex, real-world scenarios where data dimensionality poses a major challenge to model efficiency and interpretability.