Evolving Temporal Knowledge Graphs by Iterative Spatio-Temporal Walks

Hao Tang, Donghong Liu, Xinhai Xu, Feng Zhang

Abstract: Predicting facts that occur in the future is a challenging task in temporal knowledge graphs (TKGs). TKGs represent temporal facts about entities and their relations, where each fact is associated with a timestamp. Inspired from the human inference process that predictions are usually made by analyzing relevant historical clues, in this paper, we propose a model based on temporal evolution and temporal graph attention mechanism to infer future facts. Specifically, we construct a node pool to keep the importance of all nodes encountered in the historical search. We learn temporal evolution features and sub-graph structures based on temporal random walks and graph attention networks. Moreover, these sub-graphs are sets of objects with the same subjects and relations as the query. Experiments on five temporal datasets demonstrate the effectiveness of the model compared with the state-of-the-art methods. Codes are available at https://github.com/lendie/SWGAT.