GAttNHP Improves Event Forecasting in Temporal Knowledge Graphs

This research introduces GAttNHP (Group Attention Neural Hawkes Process), a novel framework designed to significantly improve the forecasting of future events within Temporal Knowledge Graphs (TKGs). TKGs, which record evolving facts over time, present several inherent difficulties for prediction: encoding long-range temporal dependencies, modeling mutual excitation or inhibition between different event chains, and handling heavy-tailed, statistically sparse inter-arrival times that make deterministic predictions unreliable. GAttNHP addresses these three core issues through a tightly integrated architecture. First, a self-attention encoder processes each subject-relation chain as a continuous-time point process, effectively capturing the lingering influence of distant historical events. Second, a semantic soft-grouping module transforms globally learnable Hawkes priors into an analytical cross-attention mask. This allows event chains to share excitation patterns based on their latent group memberships, avoiding computationally intensive exhaustive pairwise comparisons. Finally, a Non-Crossing Quantile (NCQ) regression head replaces traditional mean-based time prediction. This component provides calibrated, monotonically ordered quantile estimates that maintain stability even under heavy-tailed inter-arrival distributions. Extensive evaluations across six benchmark TKG datasets demonstrate that GAttNHP outperforms state-of-the-art baselines in both entity and time prediction. Ablation studies further confirm that its most substantial gains occur on long-tail event chains, precisely where existing models typically struggle the most.