This Friday at the “Fairness, Accountability and Transparency in Machine Learning” workshop, Sorelle will be presenting some work she, Suresh and I have been doing this year. We’ve been studying the intersection of computational ideas and disparate impact, a legal theory of discrimination that central to US anti-discrimination law.

Why should a visualization researcher care, you say?

One of the themes of my research has been that user data (typically process provenance in form of activity logs, like the version trees in VisTrails and the git logs in RCloud) can be put to very good use. We can suggest new visualizations like we did in VisComplete, or study student dynamics in homework assignments. Understandably, when I give talks about these ideas, people ask me whether users are afraid or reluctant to share data, and my answer has so far been that “cultural problems require cultural solutions”. What I meant is if problems might arise from the mere presence of the data, then we need to make sure the root cause is eliminated (“can we just stop being jerks to other people?”). This is still true, but I have come to the realization that the problems may be getting worse because of data collection and mining.

We CS researchers ought to be doing something about this. At the very least, let’s take a good, hard look at the risks.

If our culture causes a problem, then data mining can ossify this problem just by mirroring our bad institutions. Then later, when we look at our data mining outputs, we go “hey, men must be better violinists than women, because that’s what the stats say”. Look: numbers! They can’t be biased, right? boyd and Crawford have written extensively about this: you should totally read that piece if you haven’t yet.

This is not to say that we should give up on the project of data-driven, algorithmic decision-making. “We have met the enemy, and he is us”, after all: we have good evidence that humans are bad decision-makers. On the contrary: we should directly study how to characterize bias, how it shows up in data sources, and how we go about fixing it. It’s not only cool computer science: it’s the right thing to do.

There’s already been a lot of work in the area, which I won’t try to distill into a single blog post. It might even turn out that what Sorelle, Suresh and I have to say about this is in the grand scheme of things a minor observation. But! I became interested in visualization because of how well it connects the world of human experiences with the world of computation, algorithms, and data. Our work at FATML has nothing to do with visualization, but this process of looking at the human impact of a computational phenomenon is something I feel I ought to do more of, not less.

And, in addition, I think there’s a big opportunity for vis (and more generally, for people looking at algorithmic aspects of human-computer interaction). This would be using visualization for the betterment of mankind (and really, do you need a better reason than that?), but there’s good research questions here: can visualization and interactive exploration help detect, explain, and remove structural biases in data sources? How do we even define these questions correctly?

This whole field is just getting started. We vis people know how it can be a powerful way to bring clarity to messy problems. We know how powerful visualizations when they force us to face hard truths. (Hell, really I’m just badly repeating a point Tamara has been making for at least 8 years: “concern not only for truth, but also for justice”)

So maybe we should be thinking about how visualizations can empower people to detect sources of unfairness in their data-driven processes, and how they fix these processes. Why not?