Predicting Links in Knowledge Graphs With the Canonical Correlation Analysis and Fusing Tensor Model

Tinghui Wu, Sijie Mai, Dihu Chen, Haifeng Hu

Abstract: Relation prediction in knowledge graphs is critical for uncovering missing links between entities. Previous models mostly focused on learning the distance of entities and relation within each triplet. However, they relied heavily on linear metric learning-based methods to evaluate the connections between them, which ignore high-level complex interactions. Moreover, as relations and entities convey distinctive semantic information, it is difficult to correlate them in the embedding space. To address these problems, we introduce a Canonical correlation Analysis and Fusing Tensor model (CAFT) for relation prediction. Specifically, it leverages canonical correlation analysis to correlate them in the embedding space, and applies tensor fusion (TF) to comprehensively model the high-level interactions between entities and relations. As the highly expressive TF network easily leads to high computational complexity, we also derive a fusion method based on low-rank tensors. Experiments suggest that our model outperforms state-of-the-art baselines on vast datasets, including the novel biomedical dataset PharmKG.