Heterogeneous Graph Neural Network with Hypernetworks for Knowledge Graph Embedding

Xiyang Liu, Tong Zhu, Huobin Tan, Richong Zhang

Abstract: Heterogeneous graph neural network (HGNN) has drawn considerable research attention in recent years. Knowledge graphs contain hundreds of distinct relations, showing the intrinsic property of strong heterogeneity. However, the majority of HGNNs characterize the heterogeneities by learning separate parameters for different types of nodes and edges in latent space. The number of type-related parameters will be explosively increased when HGNNs attempt to process knowledge graphs, making HGNNs only applicable for graphs with fewer edge types. In this work, to overcome such limitation, we propose a novel heterogeneous graph neural network incorporated with hypernetworks that generate the required parameters by modeling the general semantics among relations. Specifically, we exploit hypernetworks to generate relation-specific parameters of a convolution-based message function to improve the model’s performance while maintaining parameter efficiency. The empirical study on the most commonly-used knowledge base embedding datasets confirms the effectiveness and efficiency of the proposed model. Furthermore, the model parameters have been shown to be significantly reduced (from 415M to 3M on FB15k-237 and from 13M to 4M on WN18RR).