Title: Conformal prediction, a method that produces valid prediction sets under the assumption of exchangeability
Authors: Lars Carlsson - Centre for Machine Learning, Royal Holloway, University of London (United Kingdom) [presenting]
Abstract: Conformal prediction can be used as a classification method in both supervised and unsupervised settings. It guarantees validity in the predictions under the exchangeability assumption. We will see how conformal prediction works for classification problems. Any machine-learning method can, with conformal prediction, produce predictions of label sets given a preset confidence level. The confidence level in predictions directly corresponds to the fraction of erroneous predictions made by the conformal predictor. This validity property will be demonstrated in an example. Furthermore, we will look at some different domains where conformal prediction has been successfully applied.