A multi-typed multi-relational heterogeneous graph neural network model for complex networks

Networks are a prevalent relationship in real-world society, and the use of graphs and graph neural networks (GNNs) to model these networks and capture their characteristic relationships has shown tremendous development potential. Due to the inherent complexity of node and edge relationships within networks, heterogeneous graph neural networks (HGNNs) have become the preferred modeling approach. However, existing HGNNs primarily focus on heterogeneous graphs with only single relationships between nodes, which limits their ability to handle complex network graphs with multiple relational interactions. To effectively capture the complex node objects and multi-relational interactions in networks, this paper proposes a neural network model for multi-class multi-relational heterogeneous graphs (MMHGNN), consisting of three modules: structural feature encoding, weighted multi-relation path aggregation, and feature fusion. In the structural feature encoding module, MMHGNN employs the Four Color Theorem to color the graph and generates type encodings of edge relationships along paths, merging color features with path encodings to serve as structural features for different paths, thereby enhancing the distinguishability of nodes under complex relationships and structures. In the weighted multi-relation path aggregation module, MMHGNN aggregates neighbors along paths based on the number of edge relationships between nodes as weights and implements a balancing strategy to prevent excessive weights on long paths. In the feature fusion module, MMHGNN combines the structural features from the structural feature encoding with the embeddings from the relation aggregation module, leveraging a graph-level attention mechanism to fuse node features across different paths and generate the final node embeddings. Experiments conducted on real-world complex network datasets demonstrate the significant advantages of MMHGNN across multiple tasks.