Supervised Neighborhood Rough Set With Adaptive Neighborhood Granularity for Multilabel Feature Selection

Multilabel learning involves scenarios where each instance is associated with multiple labels simultaneously. Existing supervised neighborhood rough set methods typically rely on binary supervision, which fail to capture partial label consistency between samples in multilabel data. Moreover, the neighborhood radius is often manually specified, making neighborhood construction less adaptive to different data distributions. To address these issues, we propose a continuously supervised neighborhood approximation scheme for multilabel feature selection. Specifically, a label consistency function is introduced to quantify the semantic similarity between instances according to their associated labels. Based on this function, an adaptive neighborhood radius estimation strategy is developed to determine the radius automatically from the data. Subsequently, the supervised neighborhood rough set framework is further generalized to the multilabel setting, followed by an analysis of its fundamental properties. Building on these components, an integrated feature selection scheme is established. Experimental results on several benchmark multilabel datasets show that the proposed method can produce smaller yet effective feature subsets, while achieving predictive performance comparable to or better than that of existing competitive approaches, thereby demonstrating its effectiveness for multilabel feature selection.