66: Fairness-Constrained Optimal Model Averaging with High-Dimensional Sparsity Learning

Music City Center 

Model fairness issues have received considerable attention in the ML community. The popular class of GLMs such as logistic regression has seen various developments to improve fairness. Despite the progress, studies remain limited in high-dimensional settings without adequate understanding from statistical perspectives. In this work, we propose a novel fairness-constrained model averaging algorithm for GLMs that can aggregate a large number of sparse model candidates to generate asymptotically fair modeling solutions. It is shown to achieve near optimal estimation risk in combining for estimation improvement, including flexible scenarios that a true model does not exist or is otherwise non-sparse. To facilitate practical applications for the model averaging approach, we further propose a new fairness-assisted stepwise sparsity learning method to help generate potentially fair model candidates. In addition, the fairness-assisted stepwise method with model selection maintains consistency properties when the true model is among the sparse feasible candidates, showing delicate distinction of combining for estimation improvement versus adaptation.

combining for improvement

high-dimensional regression

model fairness

social equity

stepwise regression 

Business and Economic Statistics Section