Bi-directional mapping for multi-label learning of label-specific features

In multi-label learning, scholars have proposed many multi-label learning algorithms that explore label-specific features in recent years. Previous studies tend to focus only on the forward projection of the instance feature space to the category label space to learn label-specific features for multi-label classification, and only simple correlations between labels are considered; however, the loss of discriminative information in the instance space and the essential connections between labels resulting from the reduction of feature dimensionality during forward projection are usually ignored. Based on the overall consideration, in this paper, we propose a bi-directional mapping for multi-label learning of label-specific features method(BDLS). Specifically, under a unified linear model for learning label-specific features for multi-label classification, we propose a novel reconstruction loss function to compensate for the loss of discriminative information generated during forward mapping. And we also propose an effective causal learning machine to explore the intrinsic causal relationships among labels for the purpose of mining the essential connections among labels. Experimental results and analysis on several multi-label datasets validate the effectiveness of our proposed method.