Few-Shot knowledge graph completion with multi-hop aggregation and relational correlation learning

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Few-shot knowledge graph completion (FKGC) methods utilize a limited number of head-tail entity pairs to learn representations of unseen relations, thereby completing missing facts. However, most existing methods exhibit two issues: 1) These methods rely exclusively on the one-hop neighbors of entities, overlooking the richer information available from multi-hop neighbors, which hinders the accurate learning of relation representations. 2) During the relation representation learning process, existing methods presume that relations are mutually independent, disregarding their correlation, resulting in suboptimal FKGC performance. In fact, through the semantic analysis of relations on two public datasets, we find that the presumption of independence among these relations does not hold. To address these limitations, we propose a general FKGC method aggregating rich multi-hop neighbors and considering the correlation between relations. Specifically, enhanced entity representations (EER) are constructed by aggregating the multi-hop neighbor information of entities, effectively addressing the issue of insufficient contextual information resulting from the reliance on one-hop neighbors. Second, complex relation distributions are constructed to model the implicit correlations among relations, from which orthogonal relation latent vectors (ORLVs) are sampled to capture the semantic correlation between relations explicitly. Finally, a multi-head attention mechanism is employed to guide the fusion of relational correlation information from ORLVs with EER, further enhancing the accuracy of relation representations. Extensive comparative experiments on four public datasets demonstrate that our method significantly improves performance over state-of-the-art baselines.