FAI: Advancing Optimization for Threshold-Agnostic Fair AI Systems

点击翻译按钮获取中文摘要

Artificial intelligence (AI) and machine learning technologies are being used in high-stakes decision-making systems like lending decision, employment screening, and criminal justice sentencing. A new challenge arising with these AI systems is avoiding the unfairness they might introduce and that can lead to discriminatory decisions for protected classes. Most AI systems use some kinds of thresholds to make decisions. This project aims to improve fairness-aware AI technologies by formulating threshold-agnostic metrics for decision making. In particular, the research team will improve the training procedures of fairness-constrained AI models to make the model adaptive to different contexts, applicable to different applications, and subject to emerging fairness constraints. The success of this project will yield a transferable approach to improve fairness in various aspects of society by eliminating the disparate impacts and enhancing the fairness of AI systems in the hands of the decision makers. Together with AI practitioners, the researchers will integrate the techniques in this project into real-world systems such as education analytics. This project will also contribute to training future professionals in AI and machine learning and broaden this activity by including training high school students and under-represented undergraduates. This project focuses on advancing optimization for threshold-agnostic fair AI systems. The research activities include: (i) developing scalable stochastic optimization algorithms for optimizing a broad family of rank-based threshold-agnostic objectives; (ii) developing novel threshold-agnostic fairness measures including Receiver Operating Characteristic curve (ROC) fairness, Area under the ROC Curve (AUC) fairness, etc. and studying the relationship between them and the existing fairness measures; (iii) developing efficient stochastic methods for in-processing fairness-aware learning methods to directly optimize threshold-agnostic objectives subject to new threshold-agnostic fairness-ensuring constraints; and, (iv) investigating effective end-to-end deep learning framework that not only automatically learns the feature representations, but also satisfies the fairness constraints. The algorithms will be evaluated on multiple tasks, including image recognition, recommendation, spatial-temporal hazard prediction, and predicting students’ performance.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.