Computational modeling study reveals improved civil unrest forecasting across five benchmark datasets, highlighting the benefit of jointly predicting event probabilities and initiators.
Key Points
To formulate civil unrest forecasting as a parallel multi-task learning problem that simultaneously predicts event occurrence probabilities and potential initiators.
Constructed an event graph representation encoded by a graph neural module combining graph attention mechanisms and graph convolutional networks.
Fed learned event representations into task-specific parallel predictors to jointly generate occurrence probabilities and initiator forecasts.
Evaluated model performance and training efficiency across five benchmark datasets against existing event prediction baselines.
The proposed parallel learning framework achieved superior predictive performance across all five datasets compared to single-task forecasting baselines.
The multi-task framework demonstrated high computational training efficiency while supplying decision-makers with comprehensive multi-attribute unrest forecasts.