Multi-label feature selection considering label importance-weighted relevance and label-dependency redundancy

Information theory has emerged as a prominent approach for analyzing feature relevance and redundancy in multi-label feature selection. However, traditional information theory-based methods encounter two primary issues. Firstly, when evaluating feature relevance, they fail to consider the differing importance of each label within the entire label set. Secondly, when assessing feature redundancy, they overlook the varying dependencies of the selected features on the labels. To address these issues, this paper proposes a novel multi-label feature selection method that considers label importance-weighted relevance and label-dependency redundancy. Specifically, we introduce the concept of label importance weight (LIW) to measure the significance of each label within the entire label set. Based on this LIW, we define a feature relevance term called label importance-weighted relevance (LIWR). Subsequently, we leverage the uncertainty coefficient to quantify the dependence of the selected features on the labels, treating it as a weight. Building upon this weight, we establish a feature redundancy term known as label-dependency redundancy (LDR). Finally, we formulate a feature evaluation criterion called LIWR-LDR by maximizing LIWR and minimizing LDR, accompanied by the presentation of a corresponding feature selection algorithm. Extensive experiments conducted on 25 multi-label datasets demonstrate the effectiveness of LIWR-LDR.