Abstract: Algorithmic stability holds when model fitting is insensitive to small changes in the training data. It is often seen as a means to assumption-lean inference, since it has important implications for generalization, predictive inference, and other statistical problems, without requiring distributional assumptions on the data. To reap these benefits, we should not leave stability as yet-another questionable assumption, but we also should not restrict ourselves to using a handful of specific, mathematically tractable algorithms that have been shown to be stable. In this talk, we establish that bagging—averaging models trained on random subsets of data—automatically stabilizes any black-box algorithm, with finite-sample guarantees controlled by the fraction of samples used in each subset. These results extend beyond prediction to any statistical method with outputs in a Hilbert space, and to classification through a new 'inflated argmax' that adapts to model uncertainty.

This talk is based on joint work with Rina Foygel Barber and Rebecca Willett.

https://jake-soloff.github.io/