Bi-Order Graph Attention Neural Network with Neighbor Node Correlation Embedding for Link Prediction

Node classification and link prediction attracts much attention in network topology analysis. By forecasting the potential presence of a connection between two nodes in a network, link prediction has extensive applications in areas such as power grid management, forecasting, and security. Recently, Graph Neural Network (GNN) based link prediction methods have effectively leveraged deep learning techniques for link prediction tasks and have garnered significant attention. However, many GNN-based approaches focus on feature extraction from the adjacency matrix for embedding at a single scale based on the entire graph structure. In many scenarios, the presence of a link between two nodes is influenced not only by the nodes themselves but also by potential conflicting links among their neighboring nodes at various scales. Thus, incorporating the effects of links and nodes across different scales within the graph is crucial for accurate link prediction for a given node pair. Furthermore, while heuristic information is frequently used as embedding features in link prediction, its integration into GNN-based methods has been relatively limited. To address these issues, this paper proposes a Bi-order Graph Attention Auto-Encoder with Neighbor Node Correlation Embedding (BGAEN) for link prediction. BGAEN enhances graph node features by integrating heuristic information. Additionally, BGAEN incorporates a node correlation-based graph auto-encoding loss to embed heuristic correlation features into node representations. These extracted correlation of node representations are then utilized into a bi-order graph attention neural network as the encoder to capture features across multiple neighborhood scales to alleviate the conflict risk for link prediction. Experimental results show that the BGAEN approach outperforms traditional link prediction methods. The experiments also confirm the effectiveness of the multi-scale neighborhood information and node correlation embedding designed in BGAEN.