A Crossover Strategy Integrated Dung Beetle Optimization for Global Optimization Problems and Feature Selection Problems

Abstract The standard Dung Beetle Optimization (DBO) algorithm often suffers from population stagnation, rapid loss of diversity, and frequent boundary violations, which limit its performance on complex optimization tasks. To overcome these limitations, this paper pro-poses CSIDBO, an improved DBO variant integrating three complementary strategies: horizontal crossover, precise elimination, and enhanced boundary guidance. Horizontal crossover introduces multi-dimensional information mixing to sustain diversity; precise elimination removes low-quality candidates to intensify evolutionary pressure; and the enhanced boundary mechanism adjusts infeasible individuals toward the global optimum, preventing boundary-induced stagnation. The effectiveness of CSIDBO was examined on three benchmark suites—CEC2017, CEC2020, and CEC2022—and compared with ten competitive metaheuristics including GWO, WOA, and the original DBO. Results show that CSIDBO consistently delivers leading performance across 20-D and 30-D test cases, achieving an average Friedman ranking of 1.60–1.70. In CEC2017, CSIDBO increases the optimization accuracy for F1 (30-D) by 99.99% relative to GWO and WOA while retaining the same computational complexity O(T × N × D) as DBO. To further explore its applicability, CSIDBO was combined with the KNN classifier to establish the CSIDBO-KNN feature selection model. Experiments on 16 real datasets indicate that the model attains over 95% average accuracy; on the Sonar dataset, it reaches 99.83%, improving accuracy by 5.66% while selecting only 4.31 features on average—20.2% fewer than WOA. For the Tic-Tac-Toe dataset, merely 5 selected features still achieve 83.16% accuracy, reducing the feature count by 34.8% compared with DBO. Fitness curves exhibit a typical “fast descent–steady convergence” pattern, and for the lung cancer dataset, CSIDBO yields a 16.8% reduction in fitness over GWO within 100 iterations. These findings confirm that CSIDBO effectively strengthens both exploration and exploitation through coordinated strategy integration. It achieves high optimization accuracy, strong convergence stability, and competitive feature-selection performance, making it a robust tool for global optimization and machine learning preprocessing.