In a supervised learning setting, a model fitting algorithm is unstable if small perturbations to the input (the training data) can often lead to large perturbations in the output (say, predictions returned by the fitted model). Algorithmic stability is a desirable property with many important implications such as generalization and robustness, but testing the stability property empirically is known to be impossible in the setting of complex black-box models. In this work, we establish that bagging any black-box regression algorithm automatically ensures that stability holds, with no assumptions on the algorithm or the data. Furthermore, we construct a new framework for defining stability in the context of classification, and show that using bagging to estimate our uncertainty about the output label will again allow stability guarantees for any black-box model. This work is joint with Jake Soloff and Rebecca Willett.

Zoom link: https://upenn.zoom.us/j/98220304722