Negative label-Aware and correlation-Enhanced multi-Label feature selection

Xiang Li, Huimin Fu, Xiaoou Huang, Tianyi Xie, Lingfei Ren, Wanfu Gao, Yonghao Li, Xin Yang

Abstract: Feature selection is often seen as an essential step in multi-label learning. Existing embedded feature selection methods mainly rely on positive label space when processing the label space. However, their training objectives and learning processes almost entirely depend on the given positive label information, neglecting the valuable negative label. Moreover, the original feature space encompasses both strong and weak correlations. Unlike strong correlations, weak correlations may introduce noise that hampers the accurate representation of the feature space. To this end, we propose a feature selection method named Negative Label-Aware and Correlation-Enhanced Multi-label Feature Selection (NCMFS) that leverages information from negative label space and enhances information in feature space. Subsequently, the enriched feature space information is integrated into the reconstruction of the label space. In the NCMFS method, we further construct a sparsity regularization term based on the difference between the original weight matrix and the mirrored weight matrix to alleviate the interference of ambiguous labels and induce sparsity in the feature selection matrix. Extensive experiments on multiple datasets show that NCMFS is more efficient and provides a more competitive feature subset quality than state-of-the-art methods.