M2I: A Multi-scale Meta-path Information Integration Framework for Heterogeneous Graph Node Classification

A large amount of the graphs in real world are inherently heterogeneous, containing different types of nodes and relationships. Heterogeneous graph embedding is the process of capturing the complex structural information of the graph. However, The existing methods only focus on the low-order initial meta-paths provided by the dataset, while neglecting the complex relationships between high-order nodes of the heterogeneous graph. In this paper, we propose a Multi-scale Meta-path information Integration framework for heterogeneous graph node classification (M2I), consisting of two modules: node feature extraction module and cross-view contrastive learning module. The former uses GNN and Transformer to generate node features in the graph through the network schema view and metapath view respectively, while incorporating extended metapaths under the metapath view to solve the deficiency of insufficient discriminative information in a single low-order initial metapath, and the latter jointly uses the features extracted from the above two views to alleviate the over-smoothing problem. Extensive experimental results on three datasets show that our method obtains superior performance compared to state-of-the-art methods.