融合多语义视图编码的异质图神经网络

Heterogeneous graph neural network fusing multi-semantic view encoding

关键词:  异质图  元路径  节点分类  链接预测

Key words:  heterogeneous graph  metapath  node classification  link prediction

摘要:

异质图神经网络在数据挖掘、信息检索等领域得到了广泛应用。基于元路径的方法通过聚合元路径邻域信息来捕捉异质图的复合关系,但元路径的选取大多依赖先验知识,容易导致语义信息的丢失或覆盖；同时,特征聚合过程中使用注意力机制计算开销过大,而且随着网络加深或元路径变长易引发语义混淆。为了解决以上问题,提出融合多语义视图编码的异质图神经网络。首先,为目标节点类型选取固定长度的所有元路径,分别构造子图以提取相应的语义信息；采用轻量级均值聚合器获得不同元路径子图下的节点表示,并为每种类型的元路径学习特定的关系编码,与其节点表示相结合；随后,进行特征映射并融合不同语义视图下的节点特征,获得最终表示并应用于下游任务。在5个真实数据集上进行实验,结果表明,提出的模型能够更有效地捕捉异质图中的语义信息,提高节点表示性能,在节点分类和链接预测任务中的表现在大多数情况下优于主流基线模型。结合消融实验和参数灵敏度分析进一步验证了模型的有效性。

Abstract:

Heterogeneous graph neural networks have been extensively applied in data mining, information retrieval, and related domains. The metapath-based approach captures composite relationships in heterogeneous graphs by aggregating metapath neighborhood information. However, the selection of metapaths predominantly relies on prior knowledge, which may lead to the loss or overwriting of semantic information. Additionally, the use of attention mechanisms in feature aggregation incurs substantial computational overhead, and semantic confusion may arise as the network deepens or metapaths lengthen. To address these issues, a heterogeneous graph neural network that integrates multi-semantic view encoding is proposed. Firstly, all metapaths of fixed length are selected for the target node type, and subgraphs are constructed to extract corresponding semantic information. A lightweight mean aggregator is employed to obtain node representation under different metapath subgraphs, and specific relation encodings are learned for each type of metapath to combine with node representation. Subsequently, feature mapping is carried out and node features from different semantic views are fused to derive the final representation, which is applied to downstream tasks. Experiments conducted on five real-world datasets demonstrate that the proposed model more effectively captures semantic information in heterogeneous graphs, enhances node representation performance, and outperforms mainstream baseline models in node classification and link prediction tasks in most cases. The effectiveness of the model is further validated through ablation studies and parameter sensitivity analyses.