Defensive Boosting for Online Probabilistic Forecasting
The paper proposes one online forecasting method that combines two boosting guarantees previously handled separately.
Noarov and Roth’s Defensive Booster forecasts binary outcomes chosen by an adaptive adversary. It competes in Brier score with the best predictor from the span of a weak hypothesis class, while also matching online classification-boosting guarantees when the smooth weak-learning condition holds. The method uses one weak-class learner rather than large ensembles, and the authors report strong performance with much faster runtime on synthetic and real data streams. ArXiv · AI/CL/LG's note
Noarov and Roth’s Defensive Booster forecasts binary outcomes chosen by an adaptive adversary. It competes in Brier score with the best predictor from the span of a weak hypothesis class, while also matching online classification-boosting guarantees when the smooth weak-learning condition holds. The method uses one weak-class learner rather than large ensembles, and the authors report strong performance with much faster runtime on synthetic and real data streams. ArXiv · AI/CL/LG's note
score 4