Multilevel structure fusion for community detection in heterogeneous graphs via disrupting heterophily

Existing community detection methods of heterogeneous graphs focus on homophilous heterogeneous graphs, which generally use GNNs and HGNNs. In heterogeneous graphs with heterophily, the two nodes connected by metapaths often have different labels, and existing methods result in representation collapse among nodes from different communities. To address the challenges posed by structural heterophily and to enhance the strength of homophily, an unsupervised community detection method with multilevel structure fusion oriented by homophily, called SFCDH, is proposed. This method demonstrates a strong generalization capabilities across various homophilous and heterophilous heterogeneous graphs. Specifically, SFCDH consists of two modules named hierarchical structure fusion module based on homophily ratio sampling and the unsupervised graph representation learning module based on generative and contrastive learning . The first module constructs a fine-grained high-order fusion graph based on the network homophily and divides the community into multi-granularity via a hierarchical abstraction tree to improve the generalization ability of the model to different metapath subgraphs. Subsequently, the second module uses contrastive loss based on community prototypes to guide the graph in generating dynamic masks and enhances the ability of SFCDH to learn attributes and complex semantics by using graph autoencoder . Ultimately, node embeddings at the metapath level and the tree level are combined to achieve multilevel structural fusion for unsupervised community detection. Extensive experiments on six publicly available datasets demonstrate the superiority of SFCDH over state-of-the-art baselines. Notably, in the heterophilous heterogeneous graph IMDB, SFCDH outperforms the second-place model by 11% in accuracy, which further validating its effectiveness. • Fuse metapath subgraphs with different homophily ratios for high-order graph. • Construct an abstraction tree based on the high-order graph using structural entropy. • Combine a dynamic masking strategy and contrastive loss for more discriminative masks. • Performance is improved by fusing metapath level and tree level node embeddings.