R. Jiménez Llamas, E. CARRIZOSA, P. Ramirez Cobo
Fairness in statistical and machine learning aims to avoid discriminatory solutions inherited from biases in datasets regarding to sensitive variables. In this work we undertake the fairness issue in a linear regression setting by means of Bayesian tools. In particular, under the Normal-Gamma model our method forces fair solutions by selecting the hyper-parameters in a way analogous to Empirical Bayes. The considered fairness measure, the average difference between sensitive and non-sensitive populations, avoids the need for individual information regarding the sensitive classes, which may be unavailable for privacy issues. As numerical illustrations show, the method provides a solution that balances between the fairness degree and an adequate fit to the data.
Palabras clave: Fairness, Bayesian linear models, Normal-Gamma probability distribution, Empirical Bayes, Gram–Schmidt process.
Programado
GT21 Inferencia Bayesiana I
7 de junio de 2022 12:00
Sala Audiovisuales
