Quote:
Originally Posted by taxwonk
Adder, I'm not accusing you of doing anything but buying into a myth that has been bought into my millions. The myth is that you can divide people or properties into different cohorts without taking race or socioeconomic status into account.
As an example, you build an algorithm that predicts risk of foreclosure. How are you going to divide up the properties? By zip code? By average home value? By number of foreclosures in a 2-mile radius? I defy you to pick a criteria that isn't reflective of the differences in our society.
Any argument that the algorithm is going to protect against downside and not perpetuate the race-wealth-education gaps is going to fall to the fact that an algorithm by definition has to categorize. If race, money, and education are all concentrated in a given area, then the distinguishing characteristic is reflected in the algorithm.
If you argue the algorithm serves a valuable purpose, then what is that purpose if not to exclude the riskier actors? Show me an algorithm that gets around this problem, I will concede. Any algorithm I can think of is going to perpetuate the discrimination. Anyone who uses that algorithm knows it will perpetuate the discrimination.
In short, it doesn't matter how much lipstick you apply. A pig is still a pig. I'm not getting personal with you. I'm suggesting that your premise is fatally flawed. If that gets me the angry fist of God, I will try to find a way to live with it. I tend to fall into that risk pool anyway.
|
While I don't disagree with you, I think, as a group, we are all over the place on this.
Sebby and I originally argued about whether redlining (and certain hiring practices and other things) were the product of outright racial bias. Based on my understanding of what he was saying, he was arguing that the practices were built around very shallow data. It seemed to me he was choosing to ignore the fact that those practices were not data-driven--that they were the
specific product of racial discrimination and not a by-product of business practices that were "blind" and just happened to end up being racially discriminatory for whatever reason (
eg., this neighborhood vs. that).
Presumably Sebby moved past this argument after reading the Times article and, even though he didn't acknowledge my original point, he decided to discuss how we will continue to be biased based on his mainly correct, heuristic view of where big data is taking us all. He and Adder now seem to be arguing over whether there is inherent value in discriminating based on race.
Your point is that if you are setting up a model to take into account anything that can be linked back to the result of our collective institutional racism (
eg., this neighborhood vs. that), that such model will be inherently discriminatory. While I agree with you, that's two steps down the line.
Right now, I'd settle for business approaches that aren't based on straight up
de jure racist bullshit. Hire based on qualifications and not the sound of one's name. Give me an interest rate based on the credit of people with similar finances, not my race. Etc. Once we've tackled that, let's address the
de facto discriminatory algorithm which draws its data from how we've unfairly educated whole groups of people or confined them to depressed neighborhoods.
In short, you and Adder aren't really disagreeing on anything, I think.
TM
