Abstract: Heterogeneous Graph Neural Network (HGNN) have made significant progress in various real-world applications, particularly in handling complex graph structures with multiple node and edge types. However, existing HGNN methods still face challenges, primarily relying on manually designed meta-paths and struggling to flexibly model multi-relational semantics. To address these issues, we propose a Multi-Head Relation-Transformed Heterogeneous Graph Neural Network (MHRTHGN) model. First, we construct independent relation transformation matrices and a multi-head attention mechanism for different relations to accurately capture multi-relation information. Then, a relation-aware attention mechanism is introduced to integrate features at both the node and relation levels, enhancing the discriminability of node representations. Finally, a feedforward network is employed to further improve the semantic expression capability. Comparative experiments on the ACM, IMDB, and DBLP datasets demonstrate that MHRTHGN achieves an average accuracy improvement of 3%–5% in node classification and clustering tasks, compared to representative models such as HAN and HGT, validating the effectiveness and superiority of the proposed model.

Key words: heterogeneous graph, graph neural network, graph representation learning, multi-head attention mechanism, relation-aware

摘要： 异构图神经网络（HGNN）近年来在多种实际应用中取得了显著进展，尤其是在处理具有多种节点和边类型的复杂图结构时。然而，现有的HGNN方法仍面临一些挑战，主要包括过度依赖人工设计的元路径、难以灵活建模多关系语义等问题。为了解决这些问题，本文提出了一种基于多头关系转化的异构图神经网络（MHRTHGN）模型。首先，为不同关系构建独立的关系变换矩阵和多头注意力机制，实现对多关系信息的精细刻画；然后，引入关系感知注意力机制，在节点级和关系级两层面融合不同关系的特征表示，提升节点表示的区分性；最后，通过前馈网络进一步增强语义表达能力。分别在 ACM、IMDB、DBLP 数据集上进行了对比实验，结果表明，相较于 HAN、HGT 等代表性模型，MHRTHGN 在节点分类与节点聚类任务上的准确率平均提升 3%~5%，验证了所提模型的有效性与优越性。

关键词: 异构图, 图神经网络, 图表示学习, 多头注意力机制, 关系感知