Biostatistics Dept Seminar: Estimation of constrained statistical functionals for fair machine learning

Biostatistics Department Seminar 

Title: Estimation of constrained statistical functionals for fair machine learning

Abstract: Constrained learning has become increasingly important in machine learning, especially  in algorithmic fairness where predictive models are specifically designed to meet pre-defined fairness criteria. This work studies constrained statistical learning from a statistical functional perspective. It focuses on estimating a function-valued parameter of interest, characterized as the minimizer of a risk criterion, under constraints where one or more fairness-related real-valued parameters are set to zero or bounded. 

This talk particularly focuses on counterfactual and causal constraints, which have emerged as an important framework for quantifying fairness notions. Often, closed-form solutions exist for the optimal parameter under causal constraints, offering insight into the mechanisms that enforce fairness in predictive models. Results also suggest natural estimators for the constrained parameter, which can be derived by combining estimates of unconstrained parameters from the data-generating distribution. As a result, fair machine-learning algorithms can be implemented seamlessly alongside any statistical learning method or off-the-shelf software.

Zoom Registration

To attend this event virtually via Zoom, please use the registration link:

Register

2024-2025 Monday Seminar Series

All Biostatistics seminars are held at 12:05 p.m., via Zoom (registration required) and onsite. View all seminar information here.