TKGR-RHETNE: A New Temporal Knowledge Graph Reasoning Model via Jointly Modeling Relevant Historical Event and Temporal Neighborhood Event Context

Temporal knowledge graph reasoning (TKGR) has been of great interest for its role in enriching the naturally incomplete temporal knowledge graph (TKG) by uncovering new events from existing ones with temporal information. At present, the majority of existing TKGR methods have attained commendable performance. Nevertheless, they still suffer from several problems, specifically their limited ability to adeptly capture intricate long-term event dependencies within the context of pertinent historical events, as well as to address the occurrence of an event with insufficient historical information or be influenced by other events. To alleviate such issues, we propose a novel TKGR method named TKGR-RHETNE, which jointly models the context of relevant historical events and temporal neighborhood events. In terms of the historical event view, we introduce an encoder based on the transformer Hawkes process and self-attention mechanism to effectively capture long-term event dependencies, thus modeling the event evolution process continuously. In terms of the neighborhood event view, we propose a neighborhood aggregator to model the potential influence between events with insufficient historical information and other events, which is implemented by integrating the random walk strategy with the TKG topological structure. Comprehensive experiments on five benchmark datasets demonstrate the superior performance of our proposed model (Code is publicly available at https://github.com/wanwano/TKGR-RHETNE).