Fusing Multi-Layer Graph Attention Networks and Bidirectional Path Reasoning for Knowledge Graph Completion

To address the insufficient collaborative modeling of local structural information and global path dependencies in knowledge graph completion tasks, we propose MGAT-BiKR, a model that combines multi-layer graph attention networks with bidirectional path reasoning. The multi-layer graph attention network dynamically aggregates the multi-hop neighborhood semantics of target entities. By leveraging a multi-head attention mechanism, it adaptively assigns weights to heterogeneous neighbors, thereby resolving the limitation of traditional methods in identifying key relational contexts. Simultaneously, bidirectional path reasoning is achieved via a BiLSTM encoder combined with path attention, which explicitly captures temporal dependencies in both forward and backward directions across multi-step relational paths. Experiments on the FB15k-237, WN18RR, and NELL995 datasets show that the proposed model achieves Hits@1 scores of 0.986, 0.978, and 0.967 in link prediction tasks. It outperforms the PathCon model by 1.2%, 1.8%, and 12.1%, respectively. This approach efficiently addresses long-tail relationship reasoning in sparse knowledge graphs.