Sequence-Based Modeling for Temporal Knowledge Graph Link Prediction

Lijie Li, Wenqiang Liu, Zuobin Xiong, Ye Wang

Abstract: Currently, the majority of research in temporal knowledge graph link prediction focuses on completing missing facts. Nevertheless, the utilization of knowledge graphs to forecast future facts has garnered significant scholarly attention. The attainment of efficient future fact prediction for time-series data hinges primarily on an in-depth exploration of both past historical facts and concurrent facts in the present. Presently, the majority of research in this domain lacks an all-encompassing integration of temporal points and durations in factual features, thereby hindering the effective management of two distinct types of facts with varying chronologies and ultimately disregarding their latent influence on future facts. This paper introduces an advanced representation model - the Progressive Representation Graph Attention Network (PRGAN) - which harnesses the potential of Graph Convolutional Neural Network and Recurrent Neural Network. PRGAN aims to ameliorate the existing shortcomings and augment the efficacy of future event prediction through attention-based learning of progressive representations of entities and relations in time series. We evaluated our proposed method with five event datasets. Extensive experimentation revealed that, in comparison with other baseline models, the PRGAN model displayed remarkable performance and efficiency in temporal reasoning tasks, thereby demonstrating its outstanding superiority.