NEW in Management Science!. My coauthors and I came up with a new consequentialist approach to designing equitable algorithms. Instead of imposing fairness criteria on an algorithm (like equal false negative rates), we aim for good outcomes. More in the 🧵 below! (1/)

I’m excited to join @CouncilonCJ’s AI Task Force to think through ways that the responsible use of generative AI can advance fairness, transparency, and public safety in the criminal justice system!.

RT @CouncilonCJ: AI presents significant opportunities and formidable challenges for criminal justice. Today, we (w/@RANDCorporation) are e….

Job alert!. We're hiring a clinical (teaching-based) Assistant Professor of Applied Statistics for Social Science Research at @NYU_ASH! . Application review begins on February 10, and the position would start on September 1. Apply here:

@madisoncoots @itshenryzhu @EmmaBrunskill @5harad Ultimately, we’ll likely achieve better outcomes if we think of algorithms as *policies* — and design them in a way that aims for the specific policy goals we desire. (17/17).

Learn more in our open-access paper, “Learning to be Fair: A Consequentialist Approach to Equitable Decision Making”, with @madisoncoots, @itshenryzhu, @EmmaBrunskill, and @5harad!. (16/).

Many studies have framed fairness as a mathematical problem, proposing axioms without considering the consequences. In contrast, our approach:.- Focuses on outcomes.- Devises a computational framework for learning to be fair in an efficient and cost-effective manner.(15/).

We use data from the Santa Clara County Public Defender to show that this approach would result in higher utility:.- During the learning phase AND.- After we stop learning!. Of course, this approach applies in any resource-constrained setting, not just for rides to court!.(14/)

The framework we designed uses contextual bandits and optimization to:.- Learn how people respond to rides, and then provide rides to people who need them—even while we’re still learning.- Equitably allocate rides by modeling preferences as parameters in a convex objective.(13/).

But randomized controlled trials are costly in a couple ways. First, people who would really benefit from a ride might be excluded if they’re randomized to a control arm. Second, we might waste money on rides for people who don’t need transportation assistance. (12/).

How could we make decisions like this in the real world?. One approach would be to run a randomized controlled trial to learn how people respond to rides. We could then estimate the tradeoffs at hand, and choose an tradeoff that best reflects our preferences. (11/)

This suggests that there’s no one-size-fits-all definition of fairness. Instead, we should make decisions in a way that reflects our preference for how to make difficult tradeoffs. (In practice, one could run a survey like the above to elicit preferences from people.).(10/)

To illustrate, we asked 300 Americans how they would make this tradeoff. After explaining the problem, we let them choose their preferred outcome. Most people preferred an outcome other than demographic parity—even people in the same political party!.(9/)

Fortunately, we don’t have to follow only one of these approaches. We could instead balance between these approaches in a way that feels most fair. But “what feels fair” is ultimately a matter of personal preference, and depends on the exact tradeoff in question. (8/)

This would reduce disparities in who gets a ride. But there would be real drawbacks!.- We’d pay for longer rides to court, so.- We’d provide fewer rides overall, so.- More people would go to jail for missing court. In other words, there’s an inherent tradeoff at play. (7/)

It might seem wrong to exclude most Black and Hispanic residents like this. So what should we do?. We could instead provide a ride to an equal share of court attendees from every neighborhood. (For researchers in ML fairness, this is akin to imposing demographic parity.).(6/).

But there’s a catch! Take Boston as an example. In prioritizing cheap + short rides, imagine that we drew from people who lived close to the courthouse (like in the map). By trying to be efficient with our budget, we’d exclude many Black + Hispanic residents of Boston. (5/)

So we’d have to choose who gets a ride. A natural way to do this would be to select people who:.a) Wouldn’t attend normally, but would attend if provided a ride.b) Live close to court, so the ride is cheap. This would maximize court appearances given our budget. Seems good!.(4/).

What if we wanted to improve court attendance rates even more?. One possible initiative would be providing people with free rides to court. But with a limited budget, we wouldn’t have enough funding to give everyone a ride. (3/)

Imagine a government initiative that aims to improve court attendance rates. In other research, we’ve already shown that sending automated court date reminders can increase court appearance rates and reduce pretrial incarceration. (2/).