GAttNHP: Group Attention Neural Hawkes Process for Extrapolation Reasoning in Temporal Knowledge Graphs

Forecasting future events in Temporal Knowledge Graphs (TKGs) is challenging due to the difficulty of encoding long-range temporal dependencies, the complex mutual influence between different event chains, and the heavy-tailed, sparse nature of inter-arrival times. The authors seek a unified framework that can simultaneously predict the next event (link prediction) and its occurrence time (time estimation) while addressing these three fundamental limitations.

The authors propose the Group Attention Neural Hawkes Process (GAttNHP), which treats each subject-relation pair as a continuous-time marked point process. The framework consists of three core components:

1. Self-Attention Encoder: Encodes individual event chains to capture long-range temporal dependencies and the lingering excitation of historical events.
2. Semantic Soft-Grouping: Dynamically assigns chains to latent groups and uses a globally learnable Hawkes prior to create an analytical cross-attention mask. This allows the model to capture mutual excitation between chains efficiently without exhaustive pairwise computation.
3. Non-Crossing Quantile (NCQ) Regression: Replaces standard mean-based time prediction with a quantile regression head that guarantees monotonic, calibrated time estimates, providing stability against heavy-tailed inter-arrival distributions.

GAttNHP consistently outperforms state-of-the-art baselines on six benchmark TKG datasets for both future link prediction and occurrence-time estimation. Ablation studies demonstrate that the model's performance gains are most pronounced on long-tail event chains, where traditional snapshot-based models typically struggle. The NCQ head effectively provides robust point estimates and uncertainty quantification, proving superior to mean-based regression in the presence of sparse, heavy-tailed temporal data.