Risk-controlling Prediction with Distributionally Robust Optimization

TMLR Paper4922 Authors

Abstract: Conformal prediction is a popular paradigm to quantify the uncertainty of a model's output on a new batch of data. Quite differently, distributionally robust optimization aims at training a model that is robust to uncertainties in the distribution of the training data. In this paper, we examine the links between the two approaches. In particular, we show that we can learn conformal prediction intervals by distributionally robust optimization on a well chosen objective. This further entails to train a model and build conformal prediction intervals all at once, using the same data.

Submission Length: Long submission (more than 12 pages of main content)

Previous TMLR Submission Url: 4913

Changes Since Last Submission:

4913 was desk-rejected for formatting issues, this was caused by the package "fullpage" being mistakenly loaded. This is now fixed.

Assigned Action Editor: Michele Caprio

Submission Number: 4922