Feature Interactive Convolutional Network with Structure-Aware Information for Knowledge Graph Embedding

Abstract

Knowledge graph link prediction is an active research topic in knowledge base completion, that is, knowledge graph embedding. Solve the problem of incompleteness of the knowledge graph by realizing the low-dimensional embedding of entities and relations. From the beginning of the shallow and fast addition models such as TransE, TransH, and TransD, to the later DistMult, ComplEx multiplication models, and the recent ConvE based on convolutional networks, they are constantly improving. ConvE uses embedded two-dimensional convolution and multi-layer nonlinear features to model the knowledge graph. The model has achieved good results and can be extended to a larger knowledge graph. However, knowledge graph as a kind of graph structure data, ConvE does not make effective use of this. The recent Graph Convolutional Network (GCN) provides a way to learn the embedding of graph nodes by using the connectivity structure of the graph. In this article, we propose a feature interaction convolutional network with structure-aware information (SAFI), SAFI is composed of the encoder of the weighted graph convolutional network and the decoder of the convolutional network. WGCN uses the node structure, node attributes, and edge relation types of the knowledge graph, and by adding learnable weights, it is used to adjust the amount of information from neighbor nodes during local aggregation, so as to achieve more accurate graph node embedding. The decoder is based on ConvE by introducing three operations: random permutation, chequer reshaping, and circular convolution, which increases the feature interaction capabilities of the model. We have proved the feasibility and effectiveness of SAFI on the benchmark FB15K-237 and WN18RR datasets, and it has about 13% improvement over ConvE on Hits@1, Hits@10, MR and MRR.