Supervision : Vanda LUENGO

Co-supervision : François BOUCHET-Sébastien LALLÉ

Multi-criteria analysis of algorithmic fairness in artificial intelligence for education

The main objective of this thesis is to develop a multi-criteria fairness analysis method for different types of algorithms commonly used in the EDM community based on different data sets. Indeed, recent works focus on the evaluation of a single criterion (gender, ethnicity, institution of origin...) but the diversity of the criteria studied shows the multidimensional aspect required to make decisions that are globally fair. This general problem is broken down into several complementary research questions that can be investigated: 1. Is it possible to combine different fair algorithms according to different criteria to obtain a fairer global decision? 2. Can we automatically detect biases in a dataset according to various criteria in order to recommend the collection of new data from a particular population? 3. Can we neutralize the biases of different algorithms over time, by alternating some methods to balance the biases according to different criteria?

2022-2023 Publications