Ellipsoidal conformal inference for Multi-Target Regression

Soundouss Messoudi, Sébastien Destercke, Sylvain Rousseau
Proceedings of the Eleventh Symposium on Conformal and Probabilistic Prediction with Applications, PMLR 179:294-306, 2022.

Abstract

Quantifying the uncertainty of a predictive model output is of essential importance in learning scenarios involving critical applications. As the learning task becomes more complex, so does uncertainty quantification. In this paper, we consider the task of multi-target regression and propose a method to output ellipsoidal confidence regions whose shapes are tailored to each instance to predict. We also guarantee that those confidence regions are well-calibrated, i.e., that they cover the ground truth with a specified probability. To achieve such a feat, we propose a conformal prediction method outputting ellipsoidal prediction regions. Experiments on both simulated and real-world data sets show that our methods outperform existing ones.

Cite this Paper


BibTeX
@InProceedings{pmlr-v179-messoudi22a, title = {Ellipsoidal conformal inference for Multi-Target Regression}, author = {Messoudi, Soundouss and Destercke, S\'{e}bastien and Rousseau, Sylvain}, booktitle = {Proceedings of the Eleventh Symposium on Conformal and Probabilistic Prediction with Applications}, pages = {294--306}, year = {2022}, editor = {Johansson, Ulf and Boström, Henrik and An Nguyen, Khuong and Luo, Zhiyuan and Carlsson, Lars}, volume = {179}, series = {Proceedings of Machine Learning Research}, month = {24--26 Aug}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v179/messoudi22a/messoudi22a.pdf}, url = {https://proceedings.mlr.press/v179/messoudi22a.html}, abstract = {Quantifying the uncertainty of a predictive model output is of essential importance in learning scenarios involving critical applications. As the learning task becomes more complex, so does uncertainty quantification. In this paper, we consider the task of multi-target regression and propose a method to output ellipsoidal confidence regions whose shapes are tailored to each instance to predict. We also guarantee that those confidence regions are well-calibrated, i.e., that they cover the ground truth with a specified probability. To achieve such a feat, we propose a conformal prediction method outputting ellipsoidal prediction regions. Experiments on both simulated and real-world data sets show that our methods outperform existing ones. } }
Endnote
%0 Conference Paper %T Ellipsoidal conformal inference for Multi-Target Regression %A Soundouss Messoudi %A Sébastien Destercke %A Sylvain Rousseau %B Proceedings of the Eleventh Symposium on Conformal and Probabilistic Prediction with Applications %C Proceedings of Machine Learning Research %D 2022 %E Ulf Johansson %E Henrik Boström %E Khuong An Nguyen %E Zhiyuan Luo %E Lars Carlsson %F pmlr-v179-messoudi22a %I PMLR %P 294--306 %U https://proceedings.mlr.press/v179/messoudi22a.html %V 179 %X Quantifying the uncertainty of a predictive model output is of essential importance in learning scenarios involving critical applications. As the learning task becomes more complex, so does uncertainty quantification. In this paper, we consider the task of multi-target regression and propose a method to output ellipsoidal confidence regions whose shapes are tailored to each instance to predict. We also guarantee that those confidence regions are well-calibrated, i.e., that they cover the ground truth with a specified probability. To achieve such a feat, we propose a conformal prediction method outputting ellipsoidal prediction regions. Experiments on both simulated and real-world data sets show that our methods outperform existing ones.
APA
Messoudi, S., Destercke, S. & Rousseau, S.. (2022). Ellipsoidal conformal inference for Multi-Target Regression. Proceedings of the Eleventh Symposium on Conformal and Probabilistic Prediction with Applications, in Proceedings of Machine Learning Research 179:294-306 Available from https://proceedings.mlr.press/v179/messoudi22a.html.

Related Material