Contrastive Learning on Subgraph and Path Structures for Inductive Knowledge Graph Completion

Subgraph-based methods have demonstrated strong performance in inductive knowledge graph completion. However, existing approaches often treat entity features as single vectors, overlooking the complexity of entity feature dimensions. To address this, we propose a Contrastive Learning framework for subgraph and path structures (CSP-IKGC), which efficiently learns enclosing subgraph features through an alternating message-passing network. The Adaptive Graph Attention Network (AGAN) treats each feature dimension of an entity as an independent channel, overcoming traditional feature aggregation methods and significantly enhancing generalization ability in inductive learning. Additionally, the path layer aggregation mechanism introduces path information to compensate for the lack of structural data, improving feature learning and reasoning accuracy in sparse graph scenarios. Experimental results demonstrate the superiority of our model across all aspects.