CAREER: Information-Theoretic Measures for Fairness and Explainability in High-Stakes Applications

点击翻译按钮获取中文摘要

Machine learning is becoming increasingly prevalent in various aspects of our lives, including several high-stakes applications, such as finance, education, and employment. These machine learning models have shown remarkable success at learning patterns present in the historical data. However, indiscriminate learning of all patterns can sometimes lead to unintended consequences, such as perpetuating disparities based on gender, race, and other protected attributes, that can adversely affect certain groups of people. This project seeks to advance the foundations of ethical and socially-responsible machine learning by empowering users to systematically identify, explain, and mitigate the sources of disparity. Rethinking the traditional paradigm of separately addressing fairness and explainability, this research project will jointly examine fairness and explainability through a unified information-theoretic lens. Furthermore, through extensive outreach and student engagements on the social impacts of machine learning, this project aims to instill interest in mathematically-principled approaches and STEM education among undergraduate and high-school students, particularly underrepresented minority students, to spearhead the next generation of socially-responsible technology.The research project will provide a novel information-theoretic view of responsible machine learning, by leveraging a body of work in information theory called Partial Information Decomposition (PID). PID is closely tethered to the principles of Blackwell sufficiency in statistical decision theory and provides a formal way of quantifying when a random variable is “more informative” than another with respect to a target variable. Combined with estimation and optimization techniques, this project will enable us to disentangle the joint information content that several random variables share about another target variable, e.g., protected attributes such as gender, race, age, nationality, etc. Four research thrusts will be investigated: (i) Providing an information-theoretic framework for explaining sources of disparity with respect to protected attributes (gender, race, etc.); (ii) Performing systematic feature selection and representation learning with disparity control; (iii) Investigating fundamental limits with a focus on distributed and federated settings; and (iv) Validating these findings on real-world datasets in finance and education. This research will lay the foundational guiding principles for engineers and policymakers so that AI can truly bring about social good.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.