Explainable Subgraph Reasoning for Forecasting on Temporal Knowledge Graphs

Interest has been rising lately towards modeling time-evolving knowledge graphs (KGs) with the rapid growth of heterogeneous event data. Recently, graph representation learning approaches have become the dominant paradigm for link prediction on temporal KGs. However, the embedding-based approaches largely operate in a black-box fashion, lacking the ability to judge the results' reliability. This paper provides a link forecasting framework that reasons over query-dependent subgraphs of temporal KGs and jointly models the graph structures and the temporal context information. Especially, we propose a temporal relational attention mechanism and a human-mimic representations-update scheme to guide the extraction of an enclosing subgraph around the query. The subgraph is then expanded via a temporal neighborhood sampling and pruning. As a result, our approach provides explainable and human-understandable arguments to the forecasting task. We evaluate our model on four benchmark temporal knowledge graphs for the link forecasting task. While being more explainable, our model also obtains a relative improvement of up to 17.7 $\%$ on MRR compared to the previous best KG forecasting methods. We also conduct a survey, and the results show that the reasoning arguments extracted by the machine for knowledge forecasting are aligned with human understanding.