EAGER: Towards Fair Regression under Sample Selection Bias

Decision making models are ubiquitous in applications like employment, credit, and insurance. Increasingly, there are worries of inaccurate decisions or even discrimination from predictive decision models that have been trained on a collection of data. Fair machine learning has been an increasingly important topic.

Fair machine learning models aim to learn a function for a target variable while ensuring the predicted value is fair based on a given fairness criterion. Much of the existing work focuses on fair classification. This project researches fair regression where the decision such as loan amount is continuous and focuses on the scenario where the existing data for building the model have different distributions from the model's future data.

In particular, this project deals with the sample selection bias where the values for the dependent variable from the training dataset are missing. The project aims to develop a unified framework and practical solutions for achieving rigorous fairness and high accuracy of the built regression model via bias correction and optimization techniques.

The technical aims of this project are divided into three thrusts. The first thrust develops the unified framework for fair regression under sample selection bias. The framework adopts the classic Heckman model to correct bias and enforces multiple advanced fairness notions via constrained optimization.

The second thrust applies the Lagrange duality theory and develops reduction approaches to solve constrained optimization. Theoretical studies of achieving strong duality for fairness notions and research of deriving approximation techniques for efficient optimization will be conducted in this thrust. The third thrust conducts empirical evaluation of the developed framework and algorithms in terms of prediction accuracy and fairness with benchmark datasets and real applications, implements and integrates the algorithms into open source libraries for fair machine learning.

The research findings expect to advance theoretical understanding of fair regression, improve its applicability for handling sample selection bias, and help transition of fair regression algorithms to use in real systems.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.