DFF-GAT: Dual-Channel Feature Fusion with Dynamic Attention Aggregation for Link Prediction

To address the challenges of complex relationship modeling and feature aggregation efficiency in knowledge graph link prediction tasks, this paper proposes a dual-channel feature fusion graph neural network model, DFF-GAT. The model innovatively designs the Dual-Channel Feature Fusion (DFF) mechanism, which enhances the synergy between local features and global semantics by splitting the node self-loop features and concatenating them with the message-passing features based on graph embedding algorithms. Additionally, it adopts the Graph Attention Network (GAT) as a dynamic aggregation function, utilizing its learnable attention weights to replace traditional static aggregation methods, effectively capturing the heterogeneous relationships between neighboring nodes. Experiments on the FB15K-237 and WN18RR datasets show that the model achieves HITS@1 scores of 39.3% and 59.3% for transductive link prediction tasks, respectively, and a 7.2% improvement in inductive relation reasoning tasks compared to baseline models. Ablation experiments further validate the effectiveness of the DFF mechanism and GAT dynamic aggregation, demonstrating their advantages in complex relationship modeling and feature fusion.