DS-ARO: a multi-strategy improved artificial rabbits optimization algorithm for global optimization and corporate bankruptcy prediction

To address the limitations of the Artificial Rabbits Optimization (ARO) algorithm in solving complex high-dimensional and multimodal optimization problems—such as insufficient convergence guidance, weak balance between exploration and exploitation, and lack of adaptive search control—this paper proposes a novel dual-strategy enhanced ARO algorithm, termed DS-ARO. The proposed method integrates three complementary mechanisms: (1) a convergence–diversity balanced mutation strategy based on current-to-pbest/1 to simultaneously enhance convergence pressure and population diversity; (2) an adaptive elite-guided search mechanism to improve the reliability of search directions and convergence stability through elite learning and adaptive perturbation control; and (3) a success-rate-based strategy selection mechanism to dynamically adjust the usage probability of different search strategies according to real-time optimization performance. To comprehensively evaluate the effectiveness of the proposed algorithm, extensive experiments are conducted on the CEC2017 (30-dimensional) and CEC2022 (10- and 20-dimensional) benchmark test suites, and compared with several state-of-the-art metaheuristic algorithms. Statistical analyses, including the Friedman test and Wilcoxon signed-rank test, are further employed to verify the significance of performance improvements. In addition, ablation studies are performed to quantify the contribution of each proposed strategy. The experimental results demonstrate that DS-ARO consistently achieves superior optimization accuracy, faster convergence speed, and stronger robustness compared with the competing algorithms. Furthermore, the proposed DS-ARO is applied to the hyperparameter optimization of the K-Nearest Neighbors (KNN) model for corporate bankruptcy prediction. Experimental results on the Wieslaw financial dataset show that the DS-ARO-KNN model outperforms traditional KNN and several mainstream machine learning models, achieving higher accuracy, precision, recall, and F1-score, while effectively reducing misclassification risks.