Variable-k Neighborhood Rough Set-Based Feature Reduction Algorithm

To address the problem that traditional neighborhood rough set methods have difficulty adapting to non-uniform data distributions, which often leads to inaccurate local approximations and unstable feature reduction results, this paper proposes a variable- k neighborhood rough set–based feature reduction method (VKNRS). The proposed method first adaptively determines the neighborhood size k for each sample according to the number of label-consistent samples within its δ- neighborhood, thereby achieving data-dependent fine-grained approximations. On this basis, a heuristic forward feature reduction algorithm based on dependency measures is further designed to select effective feature subsets. Finally, experiments are conducted on 12 UCI datasets using k-nearest neighbor and Naive Bayes classifiers. Experimental results demonstrate that VKNRS significantly outperforms representative neighborhood rough set methods in terms of average classification accuracy, and statistical tests further confirm its effectiveness.