Knowledge Graph Completion Method Based on Graph Contrastive attention Network

The goal of knowledge graph completion is to use the existing triplet information in the knowledge base to infer and fill those connections that have not been established. However, most methods deal with triples independently, ignoring the unique heterogeneous structure of knowledge graph and the rich contextual information contained between nodes, which limits the comprehensive mining of triple features. To solve this problem, this paper focuses on the end-to-end knowledge graph completion task and proposes a model called Graph Contrastive Attention Network (GCAT). Through the integrated attention mechanism, GCAT can capture the characteristics of entities and their neighboring relations in the local neighborhood in parallel, and integrate and encapsulate the context information of these entities. To further optimize the representation of triplet features, GCAT introduces a subgraph-level contrastive training strategy, which aims to improve the quality of the generated entity embeddings. To verify the effectiveness of the GCAT model, we conduct an extensive experimental evaluation on the link prediction task. We demonstrate the effectiveness of the GCAT model in predicting missing links in knowledge graphs.