Predicting What Happens Next: Next-Event Training and Evaluation for Temporal Knowledge Graph Extrapolation

Temporal Knowledge Graph Extrapolation (TKGE) aims to predict future events from historical events. Existing TKGE methods typically operate in a relation-specified setting, where the relation is given and the task is to predict the missing entity. However, in many real-world applications, only a single subject entity is available, and the goal is to predict both the missing relation and object entity of a future event. We refer to this setting as relation-unspecified and define the task as Subject-conditioned Relation and Object Prediction (SROP). This mismatch motivates our work. To address it, we propose Next-Event eXtrapolation Training (NEXT), a model-agnostic method that addresses SROP by shifting the optimization objective from predicting the missing entity to predicting the missing relation and then the object entity. NEXT applies to existing TKGE models equipped with an entity-history encoder and a relation-conditioned entity decoder. For such models, NEXT adapts them to SROP by adding an MLP-based prediction head that predicts the missing relation, after which the original decoder predicts the object under that relation. NEXT further introduces a new relation-aware training objective that supplements the original entity-prediction loss with a relation prediction loss. Experiments on four datasets show that NEXT raises SROP performance on ICEWS14, ICEWS18, ICEWS0515, and GDELT from 0.08%, 0.08%, 0.20%, and 0.20% to 9.35%, 4.57%, 10.87%, and 1.21%, respectively. An ablation study under identical architectures and parameter count further shows that this improvement is attributable to the change in the optimization objective rather than to any modification of the model architecture or parameters. These experiments demonstrate that NEXT enables existing models to solve SROP without compromising their original entity prediction capability.