I spent part of last week attending the ACM Conference on Fairness, Accountability and Transparency. Sorelle, Suresh, and I presented a tutorial and a paper on some of our joint work on creating a meaningful, thoughtful way to experimentally compare the wide variety of “fair ML algorithms” out there.
There’s a video stream available of my talk, but I wanted to include here my planned script in case that’s easier for folks to follow.
In a slightly unrelated note: if you haven’t attended FAT*, I highly recommend it. FAT* is going to be in Barcelona next year, so that is already a great reason to go. More seriously, I cannot think of a better place to listen and engage on what I think is the most important conversation about automation and society happening right now.
The talk script
(One slide per item in the numbered list below.)
-
I’m Carlos, and I want to talk to you today about joint work with Evan, Sonam, Derek, Sorelle, and Suresh.
I want to tell you about a study we have conducted of a number of fairness-enhancing interventions (“fair ML algorithms”) in the literature, and what happens when you attempt to compare them.
I’m a computer scientist by training, so I want to start by thanking you all for helping create and sustain this venue. I can personally tell you FAT* has fundamentally changed the way I view the world, our potential for harm, and our responsibility for addressing it. I can’t think of a more important way to spend my time, so thank you for that.
-
So what did the world look like before FAT*?
-
Until just very recently — and I can tell you I was a part of this problem too — we were all so excited about ML, and so naive about it, that we all just wanted any excuse to build an ML system. If it had anything to do with the real world, we were all giddy with excitement. “Social impact? What’s that?”
-
Right after that, we all experienced the first push-back, and our natural reaction was: “wait, this is just math! Of course it’s objective”. And then we realized just how uneducated we were as a field, by and large, and how much we have to learn.
-
Because of many of you in the audience today, fairness, accountability and transparency is seriously in the mind of policymakers, practitioners, and our students. As we have been seeing yesterday and today, we have a long way to go, but we now have an abundance of ideas and proposals to start mitigating some of the harm we’ve created. This is great!
-
There’s no better example of this explosion than Narayanan’s great tutorial just last year at this conference, called “21 definitions of fairness and their politics”. 21 definitions! Surely one of them is the real best one, and all we have to do it is find it, right?
Sadly, that’s not true - we have impossibility theorems, from our keynote speaker, general chair, and others. And some of the fairness notions do not seem to be compatible with one another. So what to do?
-
I will show you today one effort. We think that with the explosion of scholarly work, there’s a risk that different approaches to fairness will remain isolated within research groups or schools of thought, and we think that will discourage taking a broader view. We think it is important instead to make it as simple as possible to be confronted with these alternative definitions.
-
Specifically, we built a Python library that allowed us to run an extensive comparison of existing fairness interventions, using a number of fairness measures from the literature, trained on datasets that we believe are representative of the variety of real-life situations in which fair ML algorithms might be deployed.
This is the part of the talk where I confess to you that I expected that one algorithm to be fundamentally better than all the others, and that one specific fairness measure would capture most of the issues we tended to see.
-
With apologies to danah, the truth is that it’s complicated.
Here are three surprising findings from our work:
-
Variation from within pieces of the pipeline which we expected to be irrelevant sometimes dominate. For example, a number of fair ML algorithms assume a binary protected attribute, and the way in which we convert nonbinary protected attributes turns out to make a big difference for the implementations we have studied.
-
Although there are (at least!) 21 theoretically different notions of fairness, many of the measurements seem to be correlated with one another. I want to point out here that this could mean a number of things. It could be the case that the incompatibilities don’t actually matter for the ML classifiers used in practice; it could be the case that the data for which fair ML methods tend to be important does not trigger these distinctions; or it could be that our comparison is missing pieces.
-
Finally, we have found that for many ML tasks we have studied, some algorithm can achieve some notion of fairness. But it’s almost never the case that one algorithm does the best in all measures (that’s not surprising, since they often optimize for different notions). It’s also not the case that any one algorithm does reasonably well in all datasets (that’s more surprising and worrying)
-
Let me tell you more concretely what we did.
-
We’ve used five different datasets which encompass common sources of complications, like multiple protected attributes, numerical protected attributes and class imbalances. We would love to try your dataset as well. Please come to yesterday’s tutorial!
-
We’ve compiled a number of fair ML algorithms for which we could find the source code and could build it ourselves. If your technique isn’t here, this is likely our fault, and we’d like to fix it! Please come and talk to us.
-
Finally, we’ve implemented essentially all fairness and accuracy metrics we could find from the literature and that we could think of. We have more than 21 of them!
-
Here’s what we think is the most interesting finding. This is a visualization of the correlation between different measures, combined over all algorithms and datasets we have attempted. We find a general pattern of similar behavior for a number of notions of fairness, but we also find somewhat surprising results, such as the fact that the attribute-sensitive notions of accuracy don’t seem to behave significantly different than their unweighted versions.
-
Here’s a concrete example of this correlation. We are looking at a comparison between sex-conditioned true positive rates in the x axis and what we call “sex-calibration-negative”: this is the the difference (across the protected attribute “sex”) between the rate at which the algorithm makes negative predictions. We we are showing the results of the algorithms we have run on the adult dataset, where the algorithms are all intervening on the sex protected attribute.
This particular measure pair is one for which there exist impossibility results in the literature.
You can see what I mentioned earlier: the tradeoff is clearly there, but notice how different algorithms pick different points along that curve. This is an issue because our “hyperparameter” choice here appears to have to be the actual intervention to choose, rather than a hyperparameter within an algorithm.
-
So what now? Unfortunately, our conclusions are mostly introspective, and are generally bad news for the deployment of these algorithms.
-
First, we think our research reports need to be a little more careful in our reporting. For example, we should be reporting more extensively on the ways in which we have preprocessed our data to achieve the results. Otherwise, it’s hard to make meaningful comparisons.
-
Second, we would like to understand whether it’s enough to report on one measure from each of these clusters, or whether we need to find a dataset and an algorithm for which these measures behave more independently from each other.
-
Third, it seems that it’s unrealistic to expect one algorithm to work well in a variety of settings. That means that before we go tell policymakers about which fair ML intervention to use, I think we have some homework ourselves to do to understand why these results are so different from one another.
-
Finally, we would love to help you run your future analyses more easily and more comprehensibly. If there’s something you think we are missing, please join us - we would love to grow this project together.
Thank you.