Temporal Knowledge Graph Reasoning based on Graph Convolution Network

Abstract

Temporal Knowledge Graphs Reasoning that predicts future facts has been widely explored. Previous works attempted to use linear encoding to get recurring facts that appears in temporal knowledge graphs or use graph neural network to aggregate each sequence in temporal knowledge graphs. However, these models cannot well obtain the sequence information and ignore time information for each fact. To get the time dependence when reasoning future facts, we propose a Dynamic Time-aware Graph Convolution Network (DT-GCN), which applies a relation aware GCN to get graph structure for each sequence and a time aware GCN to get time information for each fact, then uses a relation aware recurrent neural network to get sequence patterns in temporal knowledge graphs. Extensive experiments demonstrate that DT-GCN can get better time information and has better performance than baselines in the task of link prediction.