Statistics seminar: Cross-Validation: a view from stability
- Date: 6 October 2022, 12:15–23:59
- Location: Ångström Laboratory, room 64119 or via Zoom (ID 699 250 9213)
- Type: Seminar
- Lecturer: Benny Avelin (Uppsala University)
- Organiser: Matematiska institutionen
- Contact person: Rolf Larsson
Welcome to attend this seminar held by Benny Avelin (Uppsala University) with the title "Cross-Validation: a view from stability".
Attend on site or via Zoom (meeting ID: 699 250 9213)
Abstract: Cross-validation / leave one out / rotation estimation is a method used to estimate expectations of estimators given the data. Traditionally it has been used for optimal bandwidth selection in kernel methods or bin-width in histogram estimators. In the context of supervised learning, it is used to find a model with smallest test error (model selection) or to select hyperparameters. The question is, when is the cross-validation estimate an accurate estimate of the true expectation?
In this talk I will discuss a recent result together with Lauri Viitasaari, where we prove that under certain stability properties of the risk we have exponential concentration of the leave one out error.