Feature selection (FS) plays a crucial role in high-dimensional data analysis by improving model performance and reducing computational complexity. However, existing metaheuristic-based FS methods often suffer from insufficient population diversity, premature convergence, and limited capability to escape local optima, which substantially constrains their effectiveness in complex search spaces. To address these challenges, this paper proposes a novel Improved Sand Cat Swarm Optimization algorithm with multi-strategy fusion (ISCSO) for feature selection. The proposed method introduces a hybrid initialization mechanism based on the Hénon chaotic map and lens imaging reverse learning to enhance population diversity. A golden sine-based phase adjustment strategy is further incorporated to achieve a more effective balance between global exploration and local exploitation. In addition, a nonlinear adaptive weight mechanism is designed to dynamically regulate the search process, while a simulated annealing-based acceptance criterion is integrated to improve the ability to escape local optima. Comprehensive experiments are conducted on the CEC2017 benchmark suite and 18 real-world datasets from the UCI repository. The results demonstrate that ISCSO achieves superior performance over state-of-the-art algorithms, obtaining the optimal results on 82.76% of benchmark functions. In feature selection tasks, ISCSO achieves the optimal average fitness on 94.44% of datasets, reduces feature dimensionality significantly, and consistently improves classification accuracy. These findings indicate that ISCSO provides a competitive and reliable solution for high-dimensional feature selection and complex optimization problems.

To address the imbalance between global exploration and local exploitation in the secretary bird optimization algorithm (SBOA), this paper presents a multi-strategy improved version termed MSISBOA. The proposed approach incorporates optimal Latin hypercube sampling during initialization to achieve a more uniform distribution of initial solutions. In the hunting phase, an adaptive Cauchy mutation factor and a boundary strategy are integrated to refine local search precision. To reduce the risk of stagnation in local optima during later iterations, a triangular walk strategy is utilized for mutation perturbation. Furthermore, the escape phase employs a combined Tent chaotic-Gaussian mutation factor and an elite retention strategy to maintain high-quality solutions while diversifying the population. The performance of MSISBOA was evaluated using the benchmark suites released for the IEEE Congress on Evolutionary Computation (CEC), including CEC-2017 and CEC-2022, against nine other swarm intelligence algorithms, with statistical results showing that MSISBOA achieved the highest average rank. Additionally, the algorithm was applied to 18 engineering optimization problems to assess its capability in solving practical constrained tasks. Experimental results indicate that MSISBOA provides competitive convergence characteristics and solution quality across the tested scenarios.