Temporal knowledge graph link predictions with query-guided temporal representation learning

Abstract

Temporal Knowledge Graph (TKG) records real-life events using timestamped facts and is used for the TKG link prediction task which is to answer an incomplete timestamped fact called the query. Existing works predict by learning entity embeddings where they represent entities with entity-related facts guided by queries to emphasize important ones. Although they generalize well, their learning with queries is limited since they guide learning with the average query which merges all queries without considering that queries in TKG represent diverse meanings. Merging diverse queries generates a vague averaged query which will mislead embedding learning and further confuse predictions. To resolve the limitation, we propose individual-query-guided learning (IndiQ) to learn clearer embeddings which faithfully realizes the nature of TKG that its records are diverse and should be modeled individually rather than averaging. Specifically, IndiQ formulates embedding learning as a weighted sum of entity-related facts and calculates weights using queries individually following the total probability theorem. Then, with the novel formulation, IndiQ guides the learning of entity embeddings using queries individually to identify important facts. Finally, IndiQ predicts future links using learned entity embeddings. Experimental results show that we achieve better performance. Visualizations further demonstrate the effectiveness of our IndiQ.