TMPA-HC: a two-stage heterogeneous multi-population algorithm with cooperative search for high-dimensional feature selection

Feature selection is a fundamental yet challenging task in machine learning, particularly in high-dimensional settings. Although swarm intelligence and evolutionary computation methods, including ant colony optimization and grey wolf optimizer, have shown promising performance in feature selection, they still face two major limitations in high-dimensional spaces. First, the selected feature subsets often contain considerable redundancy, which negatively impacts the performance of classifiers. Second, the computational cost increases rapidly with dimensionality, leading to unsatisfactory efficiency in practical applications. In response to the above challenges, this study introduces TMPA-HC, a two-stage heterogeneous multi-population framework that employs cooperative search for high-dimensional feature selection. The proposed approach adopts a two-stage framework that integrates an initial Fisher-score-based filtering stage with a subsequent wrapper-based heterogeneous multi-population optimization stage. In the second stage, the population is divided into multiple subpopulations with distinct search roles, enabling structured exploration-exploitation behaviors. To facilitate effective collaboration, TMPA-HC incorporates several cooperative mechanisms, including elite cross-population hybridization, cyclic information transfer, and subpopulation reorganization. In addition, a success-rate–driven adaptive control strategy dynamically adjusts the search intensity of each subpopulation, while lightweight elite-guided local search and stagnation-aware restart mechanisms enhance convergence stability and robustness. Comprehensive experiments on multiple high-dimensional benchmark datasets show that TMPA-HC achieves competitive feature selection performance with consistent convergence, demonstrating its effectiveness and stability in handling high-dimensional data.