Disentangled Hierarchical Attention Graph Neural Network for Recommendation

Heterogeneous information networks (HIN) have been widely used in recommendation systems, aiming to solve how to model complex interactions between entities and data sparsity issue. Due to the excellent performance of Graph Neural Networks (GNN) in representation learning, they are applied in recommendation systems based on HIN. However, most current works focusing on HIN overlook the entanglement of latent factors originating from different aspects. Besides, most of them use meta path-based methods, which fail to consider the semantic information among the paths. In this paper, we propose a Disentangled Hierarchical Attention Graph Neural Network for Recommendation (DHARec), which applies disentangled representations for nodes in HIN. Instead of relying solely on meta paths, welever age one-hop semantic relation neighbors to aggregate representations based on hierarchical attention, including intra relation and inter relation attention. Specifically, intra relation attention is primarily used to learn the contribution of a neighbor within the same semantic relation, while inter relation attention focuses on learning the importance of different semantic relations and fusing representations from these relations with appropriate weights. Extensive experimental results on three HIN-based datasets demonstrate that our approach outperforms existing methods.