Paul Hünermund explains how causal machine learning fixes three critical AI problems: fairness, robustness, and explainability. When COVID hit, A/B tests stopped working across different contexts. Causal techniques can better protect gender and race from bias in decisions.
We talk about data driven decision making a lot on this show and it seems like causal AI might be helpful for making decisions. So have you seen any examples of this?
Yes. So the, on a high level, I think the, the areas where causalml can really contribute to current AI practice is actually the big challenges that we face in AI, which is, well, fair decision making, fairness, robust decision making, and also explainable AI. So understanding why we're taking a certain decision and not the other. And that has to do because, well, explainability is related to the why question. And whenever we ask why why questions, it immediately relates to sort of counterfactual thinking. So we're comparing two possibilities, two states of the world. And why, for example, did my headache improve? Was it because of the pill that I took in the morning? So that immediately relates to AI, excuse me, to causal AI and robustness in the sense that when we're talking about a B testing, for example, there's always the question of external validity or what the computer scientists call transportability. So we've seen that in the COVID pandemic quite a lot. We've run all these AB tests in the past. Now the marketplace has fundamentally changed because, well, a pandemic is going on. Can we still rely on this experimental knowledge that we have accumulated in the past or do we need to repeat those A B test? And currently we don't have really good tools for sort of transporting causal knowledge across, let's say, markets or time. And causalml does contribute on that front, so making decision making more robust. And then the third pillar is fairness. We all know the stories about bias in algorithmic decision making and causal AI techniques are able to incorporate, for example, protection of a protected attribute like gender and race much more clearly.
That's really interesting. I'd maybe like to get into that in a bit more detail then.
So.
Can you maybe give me an example of like one of these stories where there has been a problem using traditional techniques and talk about how causal AI might mitigate those problems?
Yeah, so my favorite example that I'm also using in when I'm teaching causal machine learning here at Copenhagen Business School is, well, the Google controversy. That's how it was described at least in on Twitter a few years back. So in a nutshell, nutshell, Google was accused of underpaying female employees and there was even an investigation by the Department of Labor in the U.S. and then two years later, in 2019, Google published a blog post together with an associated white paper describing methodology that they've investigated this in detail using large data set, lots of various variables that went into this analysis, HR data from all across the world, Google offices. And surprisingly they found exactly the opposite. They found that it's according to them male employees and especially senior male employees, so senior software engineers that were underpaid and then they took action and actually raised salaries of these employees. And if you investigate this case closely, I don't think we have the time for it here, but it's very likely that this was a case of so called Simpson's paradox, which maybe statisticians know about, but essentially it's a causal inference problem because you're holding certain variables fixed in a regression, for example, something like seniority and gender, and that can lead to this kind of paradoxical situation that a causal effect that goes in one direction actually flips around. But this is due to spurious correlations that you basically create. And there, that is one example, I think, where it went, where things went wrong. Google already tried doing a fairly sophisticated analysis, like big data regression analysis to get at something like a gender wage gap in their organization, but not having the right tools to set up this model correctly and then got completely the wrong answer out of this.
