NSF-NSERC: Fairness Fundamentals: Geometry-inspired Algorithms and Long-term Implications

It is well-known that machine learning algorithms can show bias in decision making towards certain individuals or groups with sensitive attributes such as gender and ethnicity. This bias can be due to several reasons including biases in data, algorithmic biases, and biases in human interpretation. To mitigate this effect, one needs to better understand the learning procedure, especially when popular deep learning models are employed in order to aid decisionmakers in making fair machine decisions. Moreover, in sequential decision making setups, it is important to consider the long-term impact of automated decisions on fairness. This proposal aims to tackle the fairness challenge in machine learning by focusing on the effect of different biases on a deep learning model, as well as being cognizant of the long-term effects of fairness. The project engages undergraduate students in research and has several outreach activities involving pre-college students. The goal of this project is to develop theoretical frameworks for studying how data biases impact the fairness of learning-based decision-making algorithms in both static and dynamic settings. First, the effect of data imbalances on the geometry of models learned by deep neural networks when various training loss functions are utilized will be investigated. Building on the theoretical framework, principled algorithms will be designed that provably enhance fairness in decision making. Finally, models will be developed that tackle the challenge of guaranteeing fairness in dynamic settings where group imbalance and feature distributions evolve over time depending on the decisions. While the project focus is on theoretical understanding and formalizing fairness concepts in machine learning, various numerical evaluations will be conducted to implement the developed algorithms throughout the duration of the project. 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.