A Cross-Subgraph Attention Fusion and Comparison Method for Contrastive Learning Based Knowledge Graph Completion

Knowledge Graphs (KGs) furnish reliable and structured information that is vital for several downstream applications, including information retrieval and recommendation system. However, the pervasive incompleteness inherent in KGs frequently constrains the efficacy of these applications. To mitigate this limitation, researchers have introduced the Knowledge Graph Completion (KGC) task, which aims to supplement missing facts within incomplete triples. In recent times, contrastive learning has been incorporated into the domain of KGC, yielding substantial enhancements to the discriminative power of KGC models and establishing new performance benchmarks. Nevertheless, current contrastive methodologies usually face the problems of insufficient generalization ability of sparse relations, poor understanding of importance differences towards heterogeneous relations, as well as information redundancy in single-view comparison. To overcome these challenges, this work proposes a novel cross-subgraph attention fusion and comparison method, consisting of oriented noise injection, cross-subgraph attention fusion and cross-subgraph contrastive loss. Especially, it helps to enhance the neighboring aggregation procedure as well as the comparative loss function in existing models, by fully utilizing beneficial and complementary semantic signals from different views in the given KG. Furthermore, this contribution can be regarded as a flexible and easily adaptable plug-in component, engineered for seamless compatibility with extant contrastive learning based KGC architectures.