Adaptive Information Coordination for Context-Aware International Event Forecasting with Temporal Knowledge Graphs

Context-aware international event forecasting must preserve context-specific evidence while drawing on related interactions across semantic boundaries. We propose the hierarchical adaptive fusion and separation network (HA-AFSNet), which learns this balance before temporal encoding. A history-conditioned coefficient jointly fuses entity and relation states, linking information allocation to the participant-ranking objective. The evaluated task ranks candidate participants conditional on semantic context labels derived from articles associated with the queried events. Across five seeds on three GDELT regional graphs, HA-AFSNet improves mean reciprocal rank (MRR) over SeCoGD by 9.73–11.35%. Compared with a learned static gate, the mean MRR gains are 0.0109–0.0149, with positive pointwise 95% paired Student intervals under the stated distributional assumptions. With only five seed pairs, this statistical evidence is limited and conditional on the evaluated splits. The conditioning branch accounts for 1.68–1.78% of the full model’s parameters. Complete test-period coefficient trajectories show how the allocation varies across dates and receiving contexts. These results support history-conditioned information allocation for participant ranking within the supplied event-associated context labels.