HetReGAT-FC: Heterogeneous Residual Graph Attention Network via Feature Completion

Abstract

Heterogeneous graph embedding is receiving increasing attention from researchers due to the ubiquity of heterogeneous graphs (HGs). How to effectively handle the problem of missing features in HGs has become a hot research topic in recent years. However, the existing heterogeneous graph neural networks (HGNNs) with feature completion are facing two challenges: (1) Pre-training is required; and (2) Heterogeneous information in the HG is not fully explored. To this end, we propose a Heterogeneous Residual Graph Attention Network via Feature Completion (HetReGAT-FC). Specifically, we first design a Heterogeneous Residual Graph Attention Network (HetReGAT) to learn topological information. Then, the attention mechanism is adopted to complete the missing features. Finally, HetReGAT is used to learn the final node embeddings on the feature completed heterogeneous graph. To prove the effectiveness of this work, we conduct extensive experiments on three real-world datasets and compare it with ten competitive baselines. The results demonstrate that the proposed HetReGAT-FC significantly outperforms state-of-the-art methods. The codes and data of this work are available at https://github.com/ZZY-GraphMiningLab/HetReGAT-FC.