RAGAT: Relation Aware Graph Attention Network for Knowledge Graph Completion
Xiyang Liu, Huobin Tan, Qinghong Chen, Guangyan Lin
Abstract
Knowledge graph completion (KGC) is the task of predicting missing links based on known triples for knowledge graphs. Several recent works suggest that Graph Neural Networks (GNN) that exploit graph structures achieve promising performance on KGC. These models learn information called messages from neighboring entities and relations and then aggregate messages to update central entity representations. The drawback of existing GNN based models lies in that they tend to treat relations equally and learn fixed network parameters, overlooking the distinction of each relational information. In this work, we propose a Relation Aware Graph ATtention network (RAGAT) that constructs separate message functions for different relations, which aims at exploiting the heterogeneous characteristics of knowledge graphs. Specifically, we introduce relation specific parameters to augment the expressive capability of message functions, which enables the model to extract relational information in parameter space. To validate the effect of relation aware mechanism, RAGAT is implemented with a variety of relation aware message functions. Experiments show RAGAT outperforms state-of-the-art link prediction baselines on standard FB15k-237 and WN18RR datasets.