A Novel Integrating Approach Between Graph Neural Network and Complex Representation for Link Prediction in Knowledge Graph

Abstract

Deep learning brings high results in many problems, including Link Prediction on Knowledge Graphs (KGs). Although there are many techniques to implement deep learning into KGs, Graph Neural Networks (GNNs) have recently emerged as a promising direction for representing the structure of KGs as input for a decoder. With this structural information, GNNs can help to retain more information from the original graph than conventional embeddings like TransE, TransH, RESCAL. As a result, the learning model achieves higher accuracy in predicting missing links between entities in the KG. Meanwhile, several studies have successfully demonstrated the intrinsic properties of the embedding process in complex space while keeping many binary relations (symmetric and asymmetric). Thus, this paper proposes deploying GNNs into complex space to increase the model’s predictive capability. Another issue with GNNs is that they are susceptible to over-squashing when a large amount of information propagating between nodes is compressed down to a fixed representation space. As a result, we utilize a dynamic attention mechanism to minimize the adverse effects of these factors, and experiments on benchmark datasets have indicated that our proposal achieves a significant improvement compared to baseline models on almost all standard metrics.