大阪大学　数理・データ科学教育研究センター
Center for Mathematical Modeling and Data Science,The University of Osaka

Conformal prediction with data corruption, with extension to multi-variate uncertainty quantification

WenKai Xu 先生(Assistant professor The University of Warwick, U.K)

1. ホーム
2. セミナー・集中講義
3. セミナー
4. Conformal prediction with data corruption, with extension to multi-variate uncertainty quantification

大阪大学 データ科学セミナーシリーズ 第67回

Conformal prediction with data corruption, with extension to multi-variate uncertainty quantification

WenKai Xu 先生(Assistant professor The University of Warwick, U.K)

Conformal prediction provides powerful nonparametric techniques for constructing prediction intervals or prediction sets that is data-adaptive and model-free. One key assumption for conformal set guarantee is the exchangeability of the data. With the presence of corrupted observations, exchangeability assumptions are violated. We present techniques to adapt conformal prediction techniques into the general Huber contamination setting and provide the revised coverage guarantees. In addition, extending beyond univariate residuals, we present how an ordering of multivariate residuals based on reproducing kernels help to understand uncertainty quantification in more complex scenarios.