Relation-aware multiplex heterogeneous graph neural network

Mingxia Zhao, Jiajun Yu, Suiyuan Zhang, Adele Lu Jia

Abstract: In recent years, heterogeneous graph neural networks have attracted considerable attention for their powerful graph processing capabilities and effectiveness in handling multiple types of nodes and relationships. However, limited work has been carried out on multiplex heterogeneous networks where multiple relations exist between the same pair of nodes, which is more realistic in real-world applications. Two typical approaches include meta-path-based frameworks and weighted fusions of different edge types. The former suffers from information loss from the original graph, while the latter significantly increases computational costs as the number of subgraphs grows. To address these challenges, we propose a Relation-Aware Multiplex Heterogeneous Graph Neural Network named RAMHN, which effectively captures the multiple relations that exist between the same pair of nodes. Specifically, RAMHN first constructs hybrid relation matrices by fusing relationships and then designs a unique relation representation vector for each individual relationship. Finally, it utilizes the learned relation representation vectors and hybrid relation matrices to perform graph convolution, obtaining the final node representations. Extensive experiments on five real-world and publicly available datasets demonstrate that RAMHN outperforms state-of-the-art baselines on various downstream tasks.