<p>Feature selection (FS) is a critical preprocessing step that enhances classification performance while reducing computational cost and overfitting. Among various evolutionary algorithms, the artificial bee colony (ABC) algorithm shows promising performance for FS tasks. However, existing ABC-based FS methods often suffer from duplicate individuals in the population. These duplicates waste computational resources, slow convergence, and reduce population diversity. Complex feature interactions further compound these challenges and the risk of premature convergence, making it difficult to maintain an effective balance between exploration and exploitation. To address these limitations, we propose an improved ABC algorithm called DABC, which introduces a dedicated deduplication mechanism to enhance diversity and convergence efficiency. This mechanism consists of two tightly integrated stages: (i) a novel duplicate bee identifier that marks redundant individuals based on objective-level similarity, and (ii) a differential analysis that evaluates dissimilarities between duplicate bees and the overall population, guiding their adaptive update through weighted bitwise recombination. This joint strategy significantly reduces redundant evaluations, promotes diversity, and enhances the algorithm’s capacity to escape local optima. DABC is evaluated against seven state-of-the-art evolutionary algorithms across 24 benchmark classification datasets. Experimental results demonstrate that DABC achieves superior classification accuracy, faster convergence, and more compact feature subsets. It achieves the lowest mean classification error on nine datasets and ranks second on 11 others. Wilcoxon rank-sum tests confirm its statistical superiority over six baselines. AUC analysis further highlights its discriminative power, achieving 0.994 on Breast, 0.971 on MUSK1, and AUC <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\ge\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>≥</mo> </math></EquationSource> </InlineEquation> 0.91 on multi-class datasets such as URBAN1 and Vehicle. The effectiveness of its diversity-guided deduplication process is further validated through ablation studies and population diversity analysis. Overall, DABC presents a robust and efficient solution for FS, particularly in complex, high-dimensional settings.</p>