Multi-label learning problems have become a key topic in machine learning research in recent years. However, most approaches have focused on exploiting the interdependences between labels, whereas the correlations between the original features and each group of possible class labels have been rarely examined. The association degree of a selected feature is biased toward each discriminate class label. With the aim of addressing the gaps in previous studies, the current paper proposes a novel framework called multi-label learning with Relevant fEature for eAch Label. Using this mechanism, a classification model to deal with enron and medical data sets is established. The experimental results demonstrate the effectiveness and competitive performance of the proposed scheme which outperformed other multi-label classification methods significantly.