Local positive and negative label correlation analysis with label awareness for multi-label classification

In multi-label learning, exploiting label correlation, alleviating class imbalance and learning label-specific features have been hot topics to increase classification performance. In the paper, we propose a method to address the three issues simultaneously. The method, named LPLC-LA, builds a Bayesian model by exploiting the local positive and negative label correlations with label awareness. LPLC-LA consists of extracting label-specific features to obtain the local positive and negative correlation, defining two label aware weights for label imbalance and label separability, and then improving the estimation of label conditional probability through the two weights. The experimental results over eight benchmark datasets show that LPLC-LA can achieve better performance compared with other state-of-the-art approaches.