There is a paradox in a world that is increasingly dependent on AI. On the one hand, artificial intelligence can (or so we are told) make the world a better place. Algorithms, on the other hand, have no imagination or consciousness, so they only know the current state of affairs as reflected in the data they are trained on. And our current world is far from completely meritocratic and fair.
Jingyuan Yang, assistant professor of information systems and operations management at George Mason University’s Costello College of Business, suggests that this contradiction is exacerbated by traditional thinking about AI. “The standard idea is that fairness is a cost to efficiency. In the structure of traditional systems, fairness checks are assumed to have a negative impact on system performance and are added almost as an afterthought,” she says.
Is a world of “better” and optimized AI doomed to reproduce or even worsen existing inequalities? The research Yang is collaborating with Pengzhan Guo of Duke Kunshan University and Keli Xiao of Stony Brook University presents an attractive alternative. We use AI systems as a testing ground for the theoretical “fairness-performance complementarity,” or the idea that under certain conditions, fairness and performance mutually reinforce each other.
“Our ‘fairness by design’ framework leverages reinforcement learning, a type of machine learning (ML). However, unlike most machine learning algorithms, our algorithm involves multiple agents competing for finite resources in a dynamic environment rather than a static one,” Yang says. “As such, our paradigm is structurally much more similar to many real-world environments where different people compete over time for finite resources.”
Equity was integrated in two stages. First, the framework is designed to “guide” high-performing agents toward exploratory choices that can maximize reward. Yang explains: “In this framework, high-performing agents stay in exploration mode for a long time, while low-performing agents settle down to a stable path faster.” Second, options abandoned as a result of agents’ reward-seeking behavior were redistributed, with low-performing agents getting the first crack at the best opportunity.
Yang summarizes: “The exploratory activity of high-performing firms causes the system to direct opportunities to low-performing firms. In theory, this increases fairness without constraining performance while preserving individual choice.”
“Our ‘fairness by design’ framework utilizes reinforcement learning, a type of machine learning (ML). However, unlike most machine learning algorithms, our algorithm involves multiple agents competing for finite resources in a dynamic rather than a static environment. As such, our paradigm is structurally much more similar to many real-world environments where different people compete for finite resources over time.”
—Jingyuan Yang, assistant professor of information systems and operations management, Costello College of Business, George Mason University
To test the framework, researchers used a dataset consisting of detailed information on the work histories of 6.5 million professionals over a 20-year period. “Real-world data shows a high degree of inequality, with less redistribution of elite opportunities from relatively advantaged to disadvantaged employees,” Yang said.
The algorithm translated real-world job postings into opportunities offered to a hypothetical agent. The resulting career paths were analyzed from both a performance and equity perspective. Performance was defined by the sum of rewards earned by all agents over the entire period. Fairness was defined by the extent to which initial performance differences were resolved through subsequent decisions.
The results of the “Fairness by Design” framework outperformed those of eight alternative ML techniques drawn from three different methodological families, in both fairness and performance.
The researchers also adjusted the system to account for changes in people’s preferences. Early-career professionals are more likely to value their employer’s reputation and promotion potential. Later in your career, rewards related to job stability and security become more pronounced. Despite these limitations being implemented, the framework worked as intended, promoting upward mobility while improving the average quality of overall career paths.
In a follow-up study using the New York Yellow Taxi trip database, the framework was tasked with generating route recommendations for hypothetical “agents,” or taxi drivers, with varying track records. In this area, the selection set was much smaller (4,282 companies vs. 263 locations) and the time period was much shorter (2 hours instead of 20 years). Similar to the career planning example, the taxi study found that a more equitable distribution of high-quality routes increases average revenue per minute across the system.
“Since this framework has proven to be adaptable to different domains and agent preferences, we believe it can be used as a governance mechanism for different AI contexts in the future,” Yang says. Healthcare scheduling, course enrollment in higher education, and digital service delivery are some of the areas Yang is considering.
While she emphasizes that the research is still ongoing, she argues that it poses a significant challenge to standard thinking about AI. “Our formal proof establishes conditions where fairness and performance mutually reinforce each other, and our experiments show that those conditions are achievable in realistic settings. This provides a theoretical and experimental basis for our study. ” Yang says.
