Label enhancement-based feature selection for multi-label learning via adaptive neighborhood rough set

Feature selection effectively reduces dimensionality in multi-label datasets and boosts classifier performance, rendering it widely used in cutting-edge domains such as artificial intelligence. Nevertheless, due to the problem of label ambiguity, most existing multi-label feature selection algorithms cannot extract fine-grained label semantics from imprecise logical annotations. In an effort to circumvent this drawback, this paper proposes a novel multi-label feature selection framework that combines label enhancement and adaptive neighborhood rough set. Specifically, a cascade evidential k-nearest neighbor label enhancement algorithm is developed, which decouples the enhancement process into evidence collection and fusion phases. In the first phase, the reliability of neighbors is quantified through a distance-based weighting metric. Secondly, multi-source evidence is hierarchically fused using Dempster-Shafer theory to reconstruct fine-grained label distributions. Furthermore, to leverage this enriched label information, an adaptive neighborhood rough set model is established from a three-way decision perspective. It adaptively determines the neighborhood radius for each sample based on the local density distribution of the data, thereby eliminating the dependency on a static radius. This neighborhood granulation method enables a more precise capture of label-feature correlations and improves the effectiveness of feature selection. Benchmark evaluations conducted on fourteen authentic multi-label datasets indicate that our proposed method achieves significant improvements compared to seven leading-edge feature selection methods.