What If Algorithms Could Be Fair?(humanreadablemag.com) |
What If Algorithms Could Be Fair?(humanreadablemag.com) |
If members of my nation get drunk more often than some other, while it's offensive to say I am a 34% drunkard, on average it might hold; instead of forbidding this type of inference I'd rather rely on more signals to figure out what kind of person I am specifically for individualized decisions. They bypass this problem by adding "risky behavior" not contained in the input dataset so they just decide to model it as a hidden variable of Bayesian inference, where "risky behavior" might be correlated with ethnicity and red car anyway, just not visible outside. So if my nation is 34% drunkard but neighboring is only 11%, the conditional probability will likely be higher for my nation anyway, but obfuscated by the use of Bayesian hidden state. I am not sure why would that improve fairness.
It would only paralyze those who paid attention to the Toronto Declaration. You’re right because you can’t make ML fair because the universe isn’t fair, that’s a property of human judgements about facts. The facts remain the same regardless of ones feelings.
https://www.chrisstucchio.com/pubs/slides/crunchconf_2018/sl...
AI Ethics, Impossibility Theorems and Tradeoffs
This is the crux of the issue and as always, most people seem to miss it. Often “fair” is used as shorthand for “does what I think is right”.
Isn't this just a misleading way to say "holding a certain causal belief"? Why exactly would that be a bad thing? If you reject one set of causal beliefs, you necessarily hold a different set.
It's a much broader problem than that, because the direction of causation can be extraordinarily difficult to establish in general.
Changing the color of your car shouldn't change your ethnicity, but what if it does? Suppose you're white with Spanish ancestry and Hispanics are the group who like red cars. Paint your car red and some red-car-preferring Hispanics may be more inclined to associate with you and thereby cause you to be more immersed in Hispanic culture and start to identify as Hispanic rather than white.
And that's a silly one just to show that even the exemplar could be wrong. More plausibly, what if the causation between "risky behavior" and "red car" is reversed? We know that colors can affect human behavior. If getting into a red car makes you drive more aggressively then you have a direct causal chain between being more likely to buy a red car (for any reason) and being more likely to drive aggressively and get into a car crash.
That means that in order to use this you would first need to prove the direction of causation between the two behaviors. But that's a tall hill to climb when one of the factors you're trying to prove causation with is the one you don't have good data on.
There is also a straight forward way to tell when a method like this is definitely getting the math wrong -- does it make the prediction rate for that class of people worse? If your assumptions are correct then it shouldn't, so if it does then you've unambiguously failed.
And every time I express my desire for autobahns without speed restrictions to crisscross North America, whoever I'm talking to has generally been quick to inform me that Germans can have nice things like that because they are careful/skilled drivers, while Americans are reckless (wreckful) drivers and cannot be trusted at high speeds.
Now you may argue that correlation reflects causation in a particular case, sure, but in general, it is not the same, so it seems perfectly logical to me to point out that you can start building your model with certain causal assumptions and without others, without in any way disregarding your statistics.
Because, as people give lip service to constantly, but never seem to really adhere to, correlation is different from causation.
Making a system fair at the very least requires people designing the system to be fair. It's pretty clear that still does not happen, so I'm pretty skeptical of those that claim it's just around the corner.
It used to be considered fair to let people smoke when they wanted. Then it was considered fair to have smoking sections and non—smoking sections in restaurants. Now it’s considered fair to ban smoking entirely in restaurants and most public places.
You have a choice of whether or not you believe being male causes car insurance claims. That is independent of the statistical correlations. Ten times a day people say correlation is not causation, but a hundred times a day, I see people implicitly insisting that it necessarily is.
If I'm running an insurance agency and not a public policy advocacy, and my data keeps showing that men have higher accident rate than women, I can just ignore causation and build my actuarial tables based on that. I don't need a casual model here, at least not until the point I'd want to optimize my models further still, but there are diminishing returns on that.
The reason fairness has so much headway among engineers isn't just an aversion towards discrimination among educated people. It's that we all know this stuff is way jankier than we care to ever admit, and that we'd never want to be the data sausage going through the algorithm grinder.
Suppose (just as a toy example) that being young causes accidents, and the population of men is younger, but being male does not cause accidents. You are going to charge mature men too much and lose that business to a competitor with a correct causal model.
This is quite separate from the correlational data.
The insight I get from the article is that the "correctness" of your causal model can incorporate social justice or political correctness, without being objectively mistaken, because causation is not defined by measured correlations.
Humans are in reality much less fair than algorithms.
If there's an issue worth pursuing here, it's educating people to stop using average population statistics to rate individuals from populations. Usually the variance within a population makes population-level statistics useless for evaluating individuals.
Taking the example from our parallel discussion, if the data says being male is correlated with risky driving, and it seems to fit the casual model of "male -> risky", it would be wrong to reject it just on the grounds of "we're using this model to set insurance rates, so by penalizing males, the model is sexist". It may be that you can come up with a better casual model explaining the correlation - say, cultural history and path dependence - but until you can, rejecting a fitting model based on "it's unfair, reality ought not to be so" is just wrong.
Consider the case of African Americans who are discriminated against by traffic cops. Is it plausible that African Americans, in an attempt [perhaps in vain] to minimize interaction with traffic cops, are more cognizant of traffic laws and drive more conservatively than the average American? I don't know if the data supports that hypothetical, but it seems plausible to me.
Assuming that this were the case, if you were to assume that African Americans drove as well as white Americans, you would be discriminating against the African American population by failing to recognize their safer driving habits.
The point is that one is not compelled to believe in the causal link just because there is a statistical link.
So if certain causal links are politically contentious, rejecting them due to "political correctness" is completely separate from rejecting the facts, the statistics that are collected. It is political, but not in opposition to reality.
The article, as I understood it, is puncturing the assertion of objectivity by those who implicitly assert we have to regard all correlations as equally causal or else be against reason and logic.
But then you have to take the bad with the good. If it turns out that strict adherence to traffic laws that nobody else abides is actually more dangerous than following the normal flow of traffic, it would also detect that and take it into account.
You have a method for automatically deriving causal relationships from correlational data?
Though obviously if it isn't the cause then you're better off taking into account the true cause rather than only the thing that correlates with it -- which would cause the correlation with the outcome to disappear for the non-causal factor when you take into account both.
The original problem was that racists were not just taking race into account but were disproportionately penalizing certain races. They would literally just refuse to do business with black people. And then once that was made illegal, they would refuse to do business with people from black neighborhoods (redlining), i.e. use location as a proxy for race so they could continue to refuse to do business with black people under that pretext.
Normal Bayesian statistics doesn't do that because it's missing the actually racist piece of it, which is giving disproportionate weight to race (or something that correlates with race) so that you refuse disproportionately many people of a particular race for no legitimate reason.
The unfairness never came from taking into account some factor that correlates with race, or even race itself to the extent that it actually correlates with outcomes. It came from using a factor to deny service even though it didn't correlate with outcomes, or if it did then still not proportionately to the huge negative weight assigned to it. It came from giving race, or a proxy for race, disproportionate weight. Giving it proportionate weight isn't unfair, it's the only thing that is fair.