<p>Datasets used in classification problems often include numerous features, but not all contribute meaningfully to the classification task. Redundant or irrelevant features can even degrade the performance of classifiers. The selection of relevant features leads to superior model efficiency. Evolutionary computation-based feature selection techniques have recently gained a lot of popularity and are prevalent for feature selection. When data is complex, it becomes difficult for most evolutionary algorithms to explore all possible solutions effectively. To address this problem, a diversified and enhanced learning-based evolutionary particle swarm optimization algorithm (DLEPSO) is proposed in this study. A novel update strategy is proposed to overcome issues such as premature convergence and stagnation, which comprises various strategies, including comprehensive learning, Gaussian binned particles for diversified guidance, and sine-cosine-based acceleration coefficients. These strategies not only improve the diversity and convergence levels in the algorithm but also maintain a reasonable balance between them. Secondly, the differential search strategy is integrated to improve the exploitation ability of the swarm, and the particle’s personal best position is updated using an evolutionary approach. Before solving the feature selection problem, the proposed DLEPSO is tested on the CEC2017 benchmark suite, which verifies the efficacy of the proposed strategies. Finally, binary DLEPSO is compared with 10 algorithms, including state-of-the-art PSO variants and other popular methods, based on the performance of results on 22 datasets. The extensive experiments and statistical analysis using the Wilcoxon rank-sum and Friedman test, followed by post-hoc corrections, demonstrate the better search capability of the DLEPSO in reducing the number of redundant features while preserving a sufficient level of classification accuracy.</p>