Knowledge Graph Representation Learning for Link Prediction with Three-Way Decisions

Abstract

Relation prediction is one of the important tasks of knowledge graph completion, which aims to predict the missing links between entities. Although many different methods have been proposed, most of them usually follow the closed-world assumption. Specifically, these methods simply treat the unknown triples as errors, which will result in the loss of valuable information contained in the knowledge graphs (KGs). In addition, KGs exist large amounts of long-tail relations, which lack sufficient triples for training, and these relations will seriously affect inference performance. In order to address above-mentioned problems, we propose a novel relation prediction method based on three-way decisions, namely RP-TWD. In this paper, RP-TWD model first obtains the similarity between relations by K-Nearest Neighbors (KNN) to model the semantic associations between them. The semantic association between relations can be considered as supplementary information of long-tail relations, and constrain the learning of KG embeddings. Then, based on the idea of three-way decisions (TWD), the triples of specific relation are further divided into three disjoint regions, namely positive region (POS), boundary region (BND), and negative region (NEG). The introduction of BND intends to represent the uncertainty information contained in unknown triples. The experimental results show that our model has significant advantages in the task of relation prediction compared with baselines.