Collaborative Research: RI: AF: Small: Long-Term Impact of Fair Machine Learning under Strategic Individual Behavior

点击翻译按钮获取中文摘要

The development of Machine learning (ML) techniques have revolutionized society and enabled breakthroughs in various scientific fields. Despite the enormous social benefits, ML techniques have also caused ethical concerns when used to make decisions about humans. It has been evident that in many high-stakes applications, such as hiring, lending, criminal justice, and college admission, ML techniques may exhibit bias against disadvantaged or marginalized social groups or be vulnerable to individual strategic behavior. Recent studies have largely examined these as two separate issues in a static framework. However, the long-term impacts of ML techniques on the well-being of the population remain unclear. Since ML algorithms are deployed in a dynamic environment (i.e., individuals adapt their behaviors strategically and repeatedly as they interact with ML algorithms), ML developed in a static framework without considering human feedback effects may behave in an unanticipated and potentially harmful way. This project moves beyond static settings and aims to understand the long-term impacts of fair ML under dynamic human-ML interactions. Such an understanding is critical to ensure the trustworthiness of ML techniques and can be leveraged for designing effective interventions that promote long-term social welfare and equity; it may further help guide policymakers to design policies that better serve society. This project studies fairness problems in a sequential framework with humans repeatedly interacting with ML systems. Three key research questions will be addressed when investigating the long-term impacts of fair ML: (1) how to rigorously model individual strategic behavior and its impact on ML development; (2) how to validate and analyze the human behavioral model; and (3) what approaches can be taken to improve long-term human well-being? Integrating the knowledge from machine learning, stochastic control, game theory, and social sciences, this project will first establish an analytical framework that characterizes the complex sequential interactions between strategic individuals and ML. This framework could enable the rigorous analysis of the evolution of population dynamics and be further leveraged for developing effective interventions that improve social welfare and long-term equity. Finally, this project will conduct different analyses and experiments to examine the robustness and accuracy of the proposed framework and results.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.