Catégorie
DATAIA Seminars

DATAIA Seminars | «Algorithmic fairness: regression under demographic parity constraint» - Evgenii Chzhen

Bandeau image
stanislas-dehaene
Date de tri
Lieu de l'événement
Online

Share

twlkml
Chapo
As part of its scientific activities, the DATAIA Institute organises throughout the year seminars aimed at discussing about AI.
Contenu
Corps de texte

Evgenii Chzhen, postdoctoral researcher in the Probability and Statistics Team of the Laboratoire de Mathématiques d'Orsay (Université Paris-Saclay) will lead the June session of the DATAIA Seminars on: “Algorithmic fairness: regression under demographic parity constraint”.

The first part of the talk, will be focused on a general overview of algorithmic fairness. Even though we will mainly focus on so-called group fairness constraints and main approaches which are guided by those constraints, we will briefly discuss individual fairness as an appealing alternative to the group fairness.

The second part of the talk will be focussed on the regression under demographic parity constraint. Describing a connection of this setup with the problem of Wasserstein barycenters, we will propose a post-processing estimation procedure which enjoys statistical fairness and risk guarantees.

Finally, we will discuss the notion of risk/fairness trade-off in the problem of regression under the demographic parity constraint. In particular, we will propose a new unfairness measure, which allows to provide a sharp analytic description of this trade-off.

This talk is based on joint works with Ch. Denis, M Hebiri, L. Oneto, M. Pontil, and N. Schreuder.

To download the presentation, click here

To learn more : 

Ancre
Practical information
Corps de texte

The webinar will take place on 22 June at 11pm and will be live broadcasted.

It is recommended to use Google Chrome, Firefox, or the BlueJeans app (https://www.bluejeans.com/downloads) to join the webinar.

 

Don't miss out any new announcement about a new DATAIA seminar!
Subscribe to the DATAIA seminars mailing-list by clicking here