Heterogeneous graph structure learning based on feature and topology information extraction

Chao Li, Xin Li, Xiangkai Zhu, Qingtian Zeng, Hua Duan, Nengfu Xie

Abstract: Graph Structure Learning (GSL) aims to learn optimized graph structure and representation jointly. Its purpose is to enhance the performance and robustness of Graph Neural Networks (GNNs) by solving problems such as redundancy, bias, noise, incompleteness, and unreliability in graph structures. Compared with homogeneous graph, heterogeneous graph can better reflect the complex relationships in real-world scenarios. Heterogeneous Graph Neural Networks (HGNNs) are an important branch in the field of GNNs, which are designed to deal with complex graph structure data. However, the existing HGNNs are still insufficient in the selection of the multi-order neighborhood information aggregation method during the aggregation process, resulting in the problems of increased computational complexity and over-convergence of multi-layer semantics. Meanwhile, when facing rich node features and complex topology, these previous HGNNs have the problem of insufficient information utilization, which will prevent the model from fully exploiting the information and result in suboptimal graph structures and their corresponding representations being learned. To solve these problems, we propose a framework named HeterogeneousGraph Structure Learning based on Feature and Topology Information Extraction (HGSL-FTIE). Firstly, we introduce an n-hop meta-path extraction strategy. Based on it, we fuse the node features to construct the feature structural subgraph and fuse the original graph structure to construct the topology structural subgraph. Then, we creatively adopted a structural subgraph fusion method based on the More Confident Fusion (MCF) strategy to guide the fusion of the feature structural subgraph and topology structural subgraph. Finally, we can obtain the final graph structure and the corresponding node representation. We have conducted extensive experiments on three real datasets, and the experimental results show that our proposed model outperforms the existing state-of-the-art models. The data and code are available on GitHub.

External IDs:dblp:journals/mms/LiLZZDX25