Type-Augmented Link Prediction Based on Bayesian Formula

Ye Wang, Enze Luo, Lijie Li, Wenjian Tao

Abstract: Knowledge graphs (KGs) play a pivotal role in various real-world applications, but they are frequently plagued by incomplete information, which manifests in the form of missing entities. Link prediction, which aims to infer missing entities given existing facts, has been mostly addressed by maximizing the likelihood of observed triplets at the instance level. However, they ignore the semantic information most KGs contain and the prior knowledge implied by the semantic information. To address this limitation, we propose a Type-Augmented Link Prediction (TALP) approach, which builds a hierarchical feature model, computes type feature weights, trains them to be specific to different relations, encodes weights into prior probabilities and convolutional encodes instance-level information into likelihood probabilities; finally, combining them via Bayes rule to compute the posterior probabilities of entity prediction. Our proposed TALP approach achieves significantly better performance than existing methods on link prediction benchmark datasets.