CAREER: Strategic and Equity Considerations in Machine Learning

Machine learning (ML) algorithms have been adopted in a huge number of domains, helping in making not only predictions but decisions. Many of these applications do not have training and test data generated via the same process, in large part because learned models are deployed to make and inform decisions (rather than just predictions), and the data-generating process involves feedback loops.

In such “high-stakes" processes, predictions or decisions can produce very high-utility outcomes in some cases and very low-utility outcomes in others. These large differences in the utility of different outcomes creates additional concerns not generally present in classical prediction problems. First, large differences in the utility of outcomes will lead agents subject to these predictions to try and earn the highest-quality outcomes for themselves; second, system designers must consider whether their system makes similarly high-utility predictions for all demographic groups.

To understand just how well ML operates in these environments, this project is studying the performance of learning systems in the face of strategically generated data, and the extent to which high-quality predictions can be guaranteed on heterogeneous data sources, ensuring that the insights from ML will apply to many different populations rather than just the majority population. Complementing the technical research of this project are several projects aimed to broaden the communities entering into computing and machine learning specifically.

This includes the development of drop-in modules for integration into statistics classrooms, collaboratively with the University of Washington's (UW) K12 Computer Science (CS4Teachers), and a plan to evaluate the effectiveness of this approach with the UW’s Center for Evaluation & Research for STEM Equity.

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.