Machine Learning in Montpellier, Theory & Practice – Vincent Divol (ENSAE/CREST)
Algorithms are now routinely used to help make decisions in areas like hiring, school admissions, and healthcare. These tools can be helpful (they are fast and can handle lots of data) but they alsocome with risks. If the data they are trained on reflect unfair patterns from the past, the algorithms can end up repeating those same patterns. In this talk, I will introduce a way to think about and improve fairness in algorithms, called statistical fairness. I will focus on one idea in particular: demographic parity. This is the idea that an algorithm’s predictions should be spread out similarly across different groups (for example, different genders or ethnicities) regardless of other differences between them. I will present recent results on regression problems subject to a demographic parity constraint. I will focus in particular on the unaware setting, where the sensitive attribute (such as gender or race) cannot be used in decision-making—either because it is unavailable or because its use is restricted by law. Joint work with Solenne Gaucher (CMAP)
