Increasing fairness in recommender systems : a study of simple decorrelation applied to the Lecron-Fouss model

Files

Supervisors: Saerens, Marco
Faculty: Louvain School of Management
Degree label: Master [60] en sciences de gestion
Abstract: The objective of this work is to improve the fairness of a recommender system recently developed by professors F. Lecron (University of Mons) and F. Fouss (UCLouvain Mons). This method belongs to the family of collaborative filtering algorithms and has the particularity of performing a regularization involving a variance-covariance matrix computed from the products' scores. The first research question of this work is to know if it is possible to decrease the covariance/correlation between the ratings predicted by the algorithm and some so-called sensitive (or protected) attributes of the users such as their age or gender, thus avoiding or at least reducing a potential discrimination towards them. The second question is to know if such a decorrelation is not achieved at the cost of a too important deterioration of the quality of predictions. The objective is therefore to determine whether there is a good trade-off between fairness and prediction error. To do this, the original method was regularized to minimize the square of the covariance between the predictions and the sensitive attributes. Other simpler approaches using elementary decorrelation are also tried, both on the Lecron and Fouss method and on a k-NN recommender system, which will serve as a baseline. Applications of the subject of this work to management are also discussed.