Link Prediction Based on the Relational Path Inference of Triangular Structures

Xin Li, Qilong Han, Lijie Li, Ye Wang

Abstract: Link prediction is used to complete the knowledge graph. Convolutional neural network models are commonly used for link prediction tasks, but they only consider the direct relations between entity pairs, ignoring the semantic information contained in the relation paths. In addition, the embedding dimension of the relation is generally larger than that of the entity in the ConvR model, which blocks the progress of downstream tasks. If we reduce the embedding dimension of the relation, the performance will be greatly degraded. This paper proposes a convolutional model PITri-R-ConvR based on triangular structure relational inference. The model uses relational path inference to capture semantic information, while using a triangular structure to ensure the reliability and computational efficiency of relational inference. In addition, the decoder R-ConvR improves the initial embedding of the ConvR model, which solves the problems of the ConvR model and significantly improves the prediction performance. Finally, this paper conducts sufficient experiments in multiple datasets to verify the superiority of the model and the rationality of each module.