Multi-label feature selection for imbalanced data via KNN-based multi-label rough set theory
Weihua Xu, Yuzhe Li
Abstract
In the realm of multi-label feature selection, the intricacy of data structures and semantics has been escalating, rendering traditional single-label feature selection methodologies inadequate for contemporary demands to meet contemporary demands. This manuscript introduces an innovative neighborhood rough set model that integrates δ -neighborhood rough sets with k -nearest neighbor techniques, facilitating a transition from single-label to multi-label learning frameworks. The study delves into the attribute dependency concept within rough set theory and proposes a novel importance function based thereon, which can effectively quantify the significance of features within multi-label decision-making contexts. Building on this theoretical foundation, we have crafted a feature selection algorithm specifically tailored for imbalanced datasets. Extensive experiments conducted on 12 datasets, coupled with comparative analyses with 10 cutting-edge methods, have substantiated the superior performance of our algorithm in managing imbalanced datasets. This research not only offers a fresh theoretical perspective but also has significant practical implications, particularly in scenarios involving imbalanced datasets with multiple labels.