Increasingly, machine learning algorithms are being used to assist in making decisions that have the potential to critically impact real people’s lives. Take, for example, the use of algorithms to aid in deciding who is approved and who is not for a loan to purchase a house. In some cases, the use of such algorithms may be more efficient or more accurate. But the use of algorithms in these types of real-world situations also raise important questions about fairness.

As AI is given the power to affect people in their day-to-day lives, the scientists who build the algorithms have started asking themselves things like: Do algorithm decisions replicate bias in the data which comes from an imperfect society? Is the data collection method itself flawed, putting the burden of its mistakes on people who have been historically disadvantaged?

“I got interested in this area as a computer scientist,” said Lydia Liu, an assistant professor of computer science at Princeton University. She wondered, “If we know there are certain imperfections in the data, can we design algorithms that actively correct for this?”

Broadening the discourse on fairness

In a seminar on Feb. 18 held by the Center for Statistics and Machine Learning and the Center for Information Technology Policy, Liu discussed how computer scientists might view fairness in AI through the lens of  broader societal goals, rather than a technical or algorithmic issue. She further called for using an interdisciplinary framework informed by both political philosophy and empirical evaluations via real world experiments. The talk was one in a series of ongoing lunchtime seminars given by CSML participating faculty members. 

Once Liu began researching the topic of fairness in AI, it became clear that just tweaking an algorithm in order to achieve a mathematical definition of fairness wouldn’t be enough to truly address the problem. “Say, the algorithm ensures equal home loan approval rates for all sensitive groups,” said Liu. While that’s mathematically fair, it doesn’t necessarily account for disparate impact in all of its forms. “Depending on a number of factors, an algorithm like this might actually end up hurting people that were already disadvantaged.”

Collaborating with political philosopher Joshua Cohen, Liu took a cross-disciplinary approach to the questions computer scientists ask about fairness in the hopes of finding a holistic solution to the problems which arise from real-world impacts of algorithmic decision making. Her research ultimately draws on a combination of political theory and compelling real world case studies to call for computer scientists to broaden the discourse on fairness in machine learning by addressing discrimination beyond group subordination, equality of opportunity beyond organizational decision-making, and fairness beyond equality of opportunity. 

“All three of these ideas are about expanding the scope of fairness, instead of focusing just on what's fair within an algorithm and asking, ‘how do I redistribute those errors?’” said Liu. “It's about expanding our scope to include the human and institutional aspects and rethinking problem formulation.” 

Liu's work also focuses on leveraging experimental design and causal modeling to understand how algorithmic decisions affect real-world outcomes across areas like criminal justice, healthcare, and education. “Those things are traditionally evaluated with questions like, is my model good or bad based on some kind of benchmark data set?” said Liu. “But there's a rising recognition in the research community that we need better methods to evaluate algorithms in deployment and take into account how they're affecting the outcomes, not just predictions.”

As of right now, Liu is working with the National Institute for Student Success at Georgia State University on their data-driven advising system which uses analytics to help keep students on track for graduation. “The long-term goal is to be able to find out what makes the data-driven system successful and come up with a generalizable way to bring those benefits to other universities,” said Liu. 

With data-driven tools like these, there’s big potential for algorithms to impact student outcomes. Liu wants to help design these algorithms with care so ensure the real-world impacts are positive. “We rely on causal inference to understand what are the ways the advisor is helped by the algorithms and what are the ways in which some pieces of data cannot be represented algorithmically,” said Liu. “And then hopefully design a better system where the AI empowers the advisors to make a difference.”