Multi-label learning with Relief-based label-specific feature selection

Multi-label learning is an emerging paradigm exploiting samples with rich semantics. As an effective solution to multi-label learning, the strategy of label-specific features (LIFT) has been widely applied. Technically, such strategy feeds the tailored features to learning model instead of the original ones. However, tailoring features for each label may cause redundancy or irrelevance in feature space, thereby deteriorating the learning performance. To alleviate such a problem, a novel multi-label classification method named Relief-LIFT is proposed in this study. Relief-LIFT firstly leverages LIFT to generate the toiled features, and then adjusts Relief to select informative features from those toiled ones for the classification model. Experimental results on 12 real-world multi-label data sets demonstrate that, our proposed Relief-LIFT can achieve better performance as compared with other well-established multi-label classification methods.