Robust multi-label feature selection with shared label enhancement

Feature selection has attracted considerable attention due to the wide application of multi-label learning. However, previous methods do not fully consider relationship between feature sets and label but devote either them. Furthermore, conventional learning utilizes logical labels estimate relevance so that importance corresponding cannot be well reflected. Additionally, numerous irrelevant redundant degrade classification performance models. To this end, we propose a method named Robust Selection with shared Label Enhanced (RLEFS). First, obtain robust enhancement term by reconstructing from numerical imposing $$l_{2,1}$$ -norm onto term. Second, RLEFS share similar latent semantic structure matrix matrix. Third, local is considered ensure consistency information during process. Finally, integrate above terms into one joint framework, then, simple effective optimization provable convergence proposed solve RLEFS. Experimental results demonstrate superiority in comparison seven state-of-the-art methods.