DAG-HFC: Dual-domain attention and graph optimization network for heterogeneous graph feature completion

Abstract

Heterogeneous graphs consist of multiple types of edges and nodes. Existing heterogeneous graph neural networks can be understood as a node feature smoothing process guided by the graph structure, which can accurately simulate complex relationships in the real world. However, due to privacy and data scarcity in the real world, some node features are inevitably missing. In addition, with the increase of model depth and the aggregation of multiple types of meta-paths, node embeddings tend to be consistent, resulting in semantic confusion and overfitting problems. To improve the quality of node embeddings, we propose a heterogeneous graph feature completion network (DAG-HFC). DAG-HFC reconstructs the adjacency matrix through topology-based initial feature filling and Jaccard similarity, and fuses the meta-path adjacency matrix to integrate high-order semantic relations and structural information, enhancing the ability to model deep interactions between nodes. The designed Dual-domain differential multi-head attention mechanism combines global feature capture with local feature refinement to effectively alleviate semantic confusion and over-smoothing problems. We conduct extensive experiments on four heterogeneous graph datasets and show that DAG-HFC can significantly improve performance compared with other methods.