TEAM: Temporal knowledge graph reasoning based on Entity Activity and Multi-task Learning

Temporal knowledge graph reasoning aims to predict future events by leveraging historical facts. Existing research predominantly focuses on addressing fine-grained prediction at the level of event (e.g., entity prediction or relation prediction). However, some real-world applications only require attention to the frequency of entities participating in events rather than specific event details. Therefore, we introduce a new coarse-grained prediction task, entity activity prediction, which uses historical facts to predict the entity activity in future timestamps. Entity activity refers to the frequency at which an entity participates in various events under the same timestamp. We also propose a Temporal knowledge graph reasoning model based on Entity Activity and Multi-task learning (TEAM). Specifically, considering the variation of entity activities within a knowledge graph, we first design a hierarchical graph convolutional network to model the graph structure and capture the entity features effectively. Additionally, to better model relation information, we construct a hypergraph and employ hypergraph neural network to extract relation features. Finally, we adopt a multi-task learning strategy that includes entity prediction, relation prediction and entity activity prediction to train the model. Extensive experiments conducted on six public datasets demonstrate the effectiveness of the proposed model.