Link Prediction Based on Feature Mapping and Bi-Directional Convolution

Abstract

A considerable amount of research on link prediction has recently been driven by missing relationships between knowledge graph entities and the problem of the incompleteness of knowledge graphs. Some recent studies have shown that convolutional neural networks based on knowledge embeddings are highly expressive and have good performance in link prediction. However, we found that the convolutional neural network (CNN)-based models do not handle the link between relations and entities well. For this reason, this paper proposes a link prediction model (LPM) based on feature mapping and bi-directional convolution. For the modeling of the task, an encoding layer–mapping layer–decoding layer structure is used. Among these layers, the encoding layer adopts a graph attention network to encode multi-hop triad information and obtains richer encoding of entities and relationships. The mapping layer can realize the mapping transformation between entities and relations and project the entity encoding in the space of relation encoding to capture the subtle connection between entities and relations. The decoding layer adopts bidirectional convolution to merge and decode the triples in a sequential inverse order, which makes the decoding layer model more advantageous in prediction. In addition, the decoding layer also adopts the r-drop training method to effectively reduce the distribution error generated by training between models and enhance the robustness of the model. Our experiments demonstrated the effectiveness of mapping relations, bidirectional convolution, and r-drop, and the accuracy of the proposed model showed significant improvements for each evaluation metric on two datasets, WN18RR and FB15k-237.