Path-based Link Prediction on Hyper-relational Knowledge Graph

Recently, path-based Graph Neural Networks (GNNs) have achieved promising performance in the link prediction task on benchmark Knowledge Graphs (KGs). However, there is no research on leveraging path-based GNNs to promote the more general case of KGs, namely hyper-relational KGs (HKGs). To bridge this research gap, we study the path-based GNNs and discover that existing path-based GNNs fail to handle HKGs because they cannot well explore the external information (i.e., qualifiers) stored in HKGs. In this paper, we propose a novel framework, Hyper Path-based Graph Neural Network (HyperPGNN) for HKGs. Specifically, we propose a novel Hyper2Tri conversion and hyper query learner to better enable the path-based GNNs to understand qualifiers in HKG, and then capture them into the graph learning. Results show our method achieves good performance in both transductive and inductive settings. Codes are available at https://github.com/LARS-research/HyperPGNN.