Personalized Regularization Learning for Fairer Matrix Factorization

Matrix factorization is a canonical method for modeling user preferences for items. Regularization of matrix factorization models often uses a single hyperparameter tuned globally based on metrics evaluated on all data. However, due to the differences in the structure of per-user data, a globally optimal value may not be locally optimal for each individual user, leading to an unfair disparity in performance. Therefore, we propose to tune individual regularization parameters for each user. Our approach, personalized regularization learning (PRL), solves a secondary learning problem of finding the per-user regularization parameters by back-propagating through alternating least squares. Experiments on a benchmark dataset with different user group splits show that PRL outperforms existing methods in improving model performance for disadvantaged groups. We also analyze the learned parameters, finding insights into the effect of regularization on subpopulations with varying properties.