TSMS-HRO: A two-stage multi-strategy hybrid rice optimization algorithm for high-dimensional feature selection

Abstract High-dimensional feature selection remains a challenging and active topic in machine learning. Swarm intelligence and evolutionary computation have demonstrated promising results for high-dimensional feature selection, such as ant colony optimization algorithm, particle swarm optimization algorithm, and hybrid rice optimization algorithm, etc. However, these algorithms still face two major challenges: The first is the presence of excessive redundant features in the selected subset, which degrades classification performance; the second is the long runtime of existing methods, which hampers efficient search and timely solution. To address these challenges, the paper proposes a novel two-stage algorithm, termed the two-stage multi-strategy hybrid rice optimization algorithm (TSMS-HRO), specifically designed for high-dimensional feature selection. In the first stage, the minimum redundancy maximum relevance method is used to compute prior information to enhance the guidance of the feature subset search in the second stage. In the second stage, the hybrid rice optimization algorithm is enhanced through four mechanisms: enhancing the quality and diversity of the initial population with good point set and elite opposition-based learning strategies; increasing the utilization rate of maintainer line individuals with multiple adaptive differential operator selection strategies; improving the global and local search capabilities of the hybridization process with a t-distribution mutation perturbation strategy; and enhancing the flexibility and diversity of the selfing process of restorer line individuals by introducing an improved adaptive crossover strategy. To evaluate the performance of the proposed method, extensive numerical experiments were conducted using benchmark functions from CEC2022. Results are compared with other well-known algorithms, such as the whale optimization algorithm and grey wolf optimizer. Furthermore, TSMS-HRO is applied to 12 high-dimensional biomedical datasets. The experimental results show that TSMS-HRO outperforms other two-stage and metaheuristic algorithms based feature selection methods in terms of accuracy and convergence speed. For example, on the CLL_SUB_111 dataset with 11 340 dimensions, TSMS-HRO achieved an average accuracy of 95.25% with a 98.86% reduction in features, clearly surpassing other methods in both effectiveness and stability. These findings confirm that TSMS-HRO is an efficient and reliable algorithm not only for the optimization of functions with different characteristics but also for real-world optimization problems.