Link prediction algorithm based on dual-branch fusion network

Link prediction aims to infer unobserved or potentially existing edges within graph structures. It plays a crucial role in uncovering latent graph information, supplementing missing data, and discovering new knowledge. Graph Attention Networks (GATs) have emerged as a mainstream architecture for modeling complex topological relationships in link prediction, thanks to their ability to dynamically capture the importance of inter-node interactions. However, GATs still suffer from limitations such as ineffective capture of potential relationships among multi-hop neighbors and weak nonlinear fitting capability. To address these issues, this paper proposes a dualbranch link prediction network integrating graph attention and fusion attention. In the improved graph attention branch, the Kolmogorov-Arnold Network (KAN) is introduced, replacing edge weights with learnable nonlinear basis functions and parameterizing the activation function for each edge. This enhancement strengthens the network's ability to fit complex nonlinear relationships and improves its adaptability to high-dimensional graph data. The fusion attention branch combines the global average pooling of SE (Squeeze-and-Excitation) channel attention with the dual pooling operations of CAM (Channel Attention Module) channel attention. This integration leverages SE's global information compression capability and CAM's local salient feature extraction ability, significantly boosting the model's performance in complex scenarios. Experiments on public datasets Yeast, Citeseer, and ADV demonstrate that the proposed algorithm achieves AUC scores of 90.68%, 87.56%, and 88.98% respectively on 80% training data, outperforming baseline methods and effectively enhancing link prediction performance.