Dual-channel graph contrastive learning for multi-label classification with label-specific features and label correlations

In multi-label classification scenarios, the labels have both interactive correlations and their own respective characteristics. It is a meaningful but challenging task that learning the discriminative features specific to each label while simultaneously utilizing the correlations among multiple labels. Recently, the graph-based methods that can collaboratively learn label-specific features and correlated label semantics have shown tremendous potential for multi-label classification tasks. However, these approaches only calculate the co-occurrence probabilities between pairwise labels; they ignore high-order label correlations. Moreover, the topological structure of the label relational graph employed in this type of method is completely fixed, making the label correlations heavily dataset-dependent and limiting the generalization ability of the method. To address these issues, we propose a dual-channel graph contrastive learning method named DGCL to generate label-specific features using both second-order and high-order label correlations. Specifically, a hypergraph neural network is first employed to explore high-order label correlations. Second, an adaptive graph convolutional neural network is designed to model second-order label correlations. Finally, we construct a contrastive learning objective to collaboratively update these two label correlation semantics, which are subsequently used to guide the process of generating label-specific features. Experimental results on ten benchmark datasets demonstrate that the proposed method outperforms the state-of-the-art multi-label classification methods.