DFF-HGNN: Dual-Feature Fusion Heterogeneous Graph Neural Network

Abstract

Heterogeneous graph neural networks (HGNNs) have gained significant attention in deep learning due to their superior capability in processing heterogeneous graph data. However, existing HGNNs often fail to explicitly leverage relational information among nodes when utilizing the attribute information of nodes for graph representation learning, thus constraining their performance. To address this limitation, we introduce two approaches for utilizing relational information explicitly: a Relation-based Feature Enhancement Strategy (RFE-Strategy) for non-attributed heterogeneous graphs, and a Dual-Feature Fusion Heterogeneous Graph Neural Network (DFF-HGNN) for attributed heterogeneous graphs. The RFE-Strategy enhances HGNNs performance on non-attributed heterogeneous graphs through a three-step process: relational feature extraction, identity feature encoding, and feature enhancement. Meanwhile, DFF-HGNN integrates both attribute and relational features to effectively capture the heterogeneity and complexity of the graph, employing four components: separate pre-transformation, intra-type feature encoder, inter-type feature encoder, and embedding update encoder. Extensive experiments on multiple benchmark datasets demonstrate that the RFE-Strategy significantly improves the performance of HGNNs, while DFF-HGNN outperforms the state-of-the-art models.