Evo-Path: A Two-Stage Temporal Knowledge Graph Reasoning Model and Its Application in Human Behavior Prediction

Abstract

Temporal Knowledge Graphs (TKGs) contain a vast number of facts with timestamps in the real world and have more abundant semantic information compared with Knowledge Graphs (KGs). However, TKGs are usually incomplete. Reasoning potential facts in the future is a challenging task and has emerged as a hotspot in the research of TKGs. One key of this task is to dive deep into the evolutional patterns contained in historical facts, which are helpful to predict future facts. However, most of the existing models focus on modeling evolutional patterns based on the entire graph, which ignores the important role of the query-related paths. Furthermore, the existing graph neural networks (GNNs) cannot efficiently extract the graph structural information that is vital to model the evolutional pattern. In order to address these two problems, we propose a two-stage reasoning model Evo-Path in this paper. At the temporal path searching stage, Evo-Path learns a temporal-semantic policy network and employs beam search policy to obtain the clue paths and candidate answers based on Deep Reinforcement Learning (DRL). At the evolutional reasoning stage, we propose a Multi-Relational Graph Attention Network (MRGAT) to encode the structural information of clue subgraphs so as to model the evolutional patterns over clue paths and deduce final answers. We evaluate our model on four public TKG datasets: ICEWS14, ICEWS18-7000, ICEWS05-15-7000, and GDELT. Extensive experimental results show that Evo-Path not only outperforms other state-of-the-art baselines in MRR and Hits@1 but also has interpretability for reasoning results. Three application case studies highlight the significant value of our model for predicting human behavior.