Knowledge embedding via hyperbolic skipped graph convolutional networks

Abstract

With the aim of constructing a low-dimensional representation space, Hyperbolic Knowledge Embeddings have gradually become a hot spot in various information retrieval and machine learning tasks. However, most of the existing Hyperbolic knowledge embedding methods focus on the shallow embedding, and often ignore the network structure characteristics (e.g., hierarchy) of Knowledge Graphs. Therefore, this paper designs a novel Hyperbolic Skipped Knowledge Graph Convolutional Network, HSKGCN, to improve link prediction accuracy with low embedding dimension requirements. Firstly, the model is designed based on the hyperbolic geometric operations on the Poincaré ball, which can effectively utilize the characteristics of the hyperbolic geometry (e.g., Poincaré ball) to capture the hierarchy of Knowledge Graphs; Secondly, each single-layer convolutional layer introduces the feature aggregation weight, which ensures the reasonable distribution of node features during the aggregation process; In addition, the skip-connection mechanism is applied to HSKGCN to weaken the information loss caused by the stacked of the graph convolutional layers; Finally, we evaluate HSKGCN on benchmark datasets, WN18RR and FB15k-237. Experiments show that HSKGCN achieves substantial improvements against state-of-the-art models on the 32-dimensional embedding task, and the results of different relations on WN18RR show graphs similar with tree topology can performer better.