Chronological feed of everything captured from Olivier Grisel.
Competing risks survival analysis has been constrained by models that tightly couple architecture and loss function, limiting scalability. Alberge et al. introduce a strictly proper censoring-adjusted separable scoring rule that evaluates each observation independently, enabling stochastic optimization over data subsets. Built on this foundation, SurvivalBoost — a gradient boosting tree implementation — outperforms 12 state-of-the-art models on 4 real-world datasets across both competing risks and standard survival settings, while also delivering superior calibration, flexible time-horizon prediction, and faster computation.