Tensor factorization for temporal knowledge graph forecasting

Temporal knowledge graph forecasting predicts links at future, previously unseen timestamps from a history of relational events. Although tensor factorisation is effective and scalable for static knowledge graphs, forecasting work has increasingly relied on more computationally expensive recurrent, graph-neural, and transformer architectures.

This work extends TNTComplEx with a radial-basis-function timestamp encoder that can construct representations for unseen times. A temporal regulariser encourages smooth evolution in the embedding space, adapting a lightweight factorisation model to the forecasting setting.

Evaluation on five standard datasets finds that carefully designed and tuned tensor-factorisation models can match or outperform the tested deep architectures while requiring substantially less training and inference time. The proposed approach also improves strongly over previously reported factorisation baselines.

Paper: https://doi.org/10.1016/j.neucom.2026.132846