Heterogeneous Graph Neural Network with Reality Path Weight

In recent years, Heterogeneous Graph Neural Networks (HGNNs) have undergone significant development. In social network graph, the connections between two nodes can vary widely, ranging from 0 to hundreds or even thousands of interactions. Moreover, in datasets related to reality world, the types of nodes are diverse and complex, including train schedules, restaurants, shops, etc. Therefore, relying on expert systems to identify complex graph meta-paths becomes impractical. To address these challenges, this paper proposes a novel graph neural network model, which utilizes attention mechanism with reality path weight (PwHGNN). Firstly, we construct a multi-state adjacency weight matrix based on the raw data, treating this factual data as the weights for training the attention neural network. Next, we employ an automatic meta-path search method to identify meta-paths with high weight states and use the attention neural network to aggregate neighboring nodes along these meta-paths, obtaining node local structure vector representations. Finally, we aggregate node vector representations under different meta-paths based on the weights derived from the raw data. We conduct experiments on four datasets, including social network datasets and communication datasets. The experimental results demonstrate that our proposed model outperforms state-of-the-art graph neural network models, particularly on social network datasets.