Self-supervised contrastive learning on heterogeneous graphs with mutual constraints of structure and feature

Self-supervised learning on heterogeneous graphs has gained significant attention as it eliminates the need for manual labeling. However, most existing researches focus on predefined meta-paths that relies on domain knowledge, and they cannot handle the noises in graphs effectively. To address these problems, we propose a self-supervised contrastive learning method on heterogeneous graphs with mutual constraints of structure and feature called HeMuc. Specifically, in the high-order relation view, we exploit graph reachability to obtain the sequence of target nodes traversed by source nodes. Furthermore, we design degree and feature constraints to reduce the noises in topological structure. In the feature view, we reconstruct the graph structure using the similarity between node features and eliminate the dependence on the original graph. Finally, we propose a contrastive learning method by designing a new sampling strategy that combines the structure and feature information. The experimental results on the tasks of node classification and node clustering demonstrate that the proposed HeMuc outperforms the state-of-the-art methods. The source codes of this work are available at https://github.com/ZZY-GraphMiningLab/HeMuc.