Heterogeneous Graph Neural Network Model Based on Edge Feature Generation and Meta-Graph Similarity for Feature Extraction

Edges and nodes form the core elements of heterogeneous graphs (HGs). However, existing heterogeneous graph neural networks (HGNNS) largely rely on meta-paths to capture semantic information of nodes, often overlooking the features embedded in edges. This not only limits the expressiveness of the models but also the potential semantic isolation issues within meta-paths can affect the learning efficiency of the models, leading to suboptimal outcomes. To address these challenges, this paper proposes a novel heterogeneous graph neural network model—heterogeneous graph neural network model based on edge feature generation and meta-graph similarity for feature extraction (HGNN-ESFE)—which aims to optimize the feature extraction process through edge feature generation and meta-graph similarity. Given that edges in HGs typically lack direct feature descriptions, HGNN-ESFE innovatively adopts an averaging strategy to generate feature vectors for different types of edges. During this process, to ensure that edge features more comprehensively reflect the structural characteristics of the heterogeneous graph, HGNN-ESFE replaces the original features of the nodes at both ends of the edges with one-hot encoding. Subsequently, by randomly sampling a fixed number of edges from the generated edge features, HGNN-ESFE aggregates them to construct richer representations of edge features. To tackle the semantic isolation problem brought about by meta-paths, HGNN-ESFE introduces a similarity-based neighbor node screening mechanism. This mechanism effectively filters out noise information in the meta-graph, retaining only those neighbor nodes highly relevant to the target node for feature aggregation. This approach not only enhances the accuracy of feature extraction but also strengthens the model’s ability to understand complex relationships within heterogeneous graphs. Furthermore, by integrating feature information obtained from two distinct aggregation strategies, HGNN-ESFE utilizes a module-level attention mechanism to achieve effective fusion of these two parts of features, ultimately generating high-quality node embeddings. The experimental results demonstrate that, in tasks such as node classification, HGNN-ESFE achieves approximately a 1%-2% performance improvement compared to the state-of-the-art methods.