3 ms·

I feel like this is a politically-correct rant thinly disguised as programming related. > I don’t care what your politics are. If you’re building a data-driven

by PeCaN 10y ago

I feel like this is a politically-correct rant thinly disguised as programming related. > I don’t care what your politics are. If you’re building a data-driven system and you’re not actively seeking to combat prejudice, you’re building a discriminatory system. Fuck that. Build _correct_ systems first. If you’re making a system that intentionally distorts data for the sake of “combating prejudice”, you’re _lying_. That doesn’t help anyone.

jordigh 10y ago
The programmer's prejudice may show up in code. There are funny/sad examples of blacks getting identified as gorillas or not being recognised at all by face recognition software... all because the person designing the code forgot to include enough black people in the training set: http://edition.cnn.com/2009/TECH/12/22/hp.webcams/index.html http://edition.cnn.com/2009/TECH/12/22/hp.webcams/index.html http://www.businessinsider.com/google-tags-black-people-as-gorillas-2015-7 http://www.businessinsider.com/google-tags-black-people-as-g...
MereInterest 10y ago
What you call "intentionally distorting data", I would call "correcting for a confounding variable". Unrelated Example: Every morning, I weigh myself. Today, I found that my scale is inaccurate, and understates my weight by 5%. Therefore, if I am to predict my weight in the future, I must account for the inaccuracy of the scale. Related Example: There exist police statistics for arrests. Studies have found that arrests are racially biased, and therefore those police statistics do not accurately represent the crime rate, only the biased arrest rate. Therefore, if I am to predict the crime rate in the future, I must account for the inaccuracy of the data set.
- PeCaN 10y ago
  I'm not going to get into a debate if you're going to appeal to the “actual crime rate” vs the “racially biased arrest rate”. There's not enough data to know, and too many interacting variables. It's enough for me to accept that blacks represent a disproportional amount of inmates and that black neighborhoods are frequently stuck in poverty cycles. I think the solution actually has nothing to do with correcting variables for race or anything of the sort; rather, help people—regardless of race—out of poverty and the statistics will level out. I'm extremely skeptical that arrest “bias” is skin color related. Also, citation needed for “studies have found that arrests are racially biased”. That is entirely too hard to measure. Well, it's not that I doubt such studies exist, I simply doubt they're valid in any way, and there are probably other studies with the opposite conclusion.
- programmarchy 10y ago
  But the problem you describe isn't biased data, it's state law enforcement agencies practicing discrimination. Seems like it'd be more effective to strike at the root, not the branches.
  - MereInterest 10y ago
    Whether or not data are biased depends on what you want to measure. If you are trying to measure the arrest rate, then the arrest rates are an unbiased predictor. If you are trying to measure the true crime rate, then the arrest rates are a biased predictor. I certainly agree that racially motivated arrests are a large issue, and should be dealt with. Avoiding misinterpreting data is a widespread issue that in these cases happens to have prejudicial consequences.

by PeCaN 10y ago

jordigh 10y ago

The programmer's prejudice may show up in code. There are funny/sad examples of blacks getting identified as gorillas or not being recognised at all by face recognition software... all because the person designing the code forgot to include enough black people in the training set: http://edition.cnn.com/2009/TECH/12/22/hp.webcams/index.html http://edition.cnn.com/2009/TECH/12/22/hp.webcams/index.html http://www.businessinsider.com/google-tags-black-people-as-gorillas-2015-7 http://www.businessinsider.com/google-tags-black-people-as-g...

MereInterest 10y ago

What you call "intentionally distorting data", I would call "correcting for a confounding variable". Unrelated Example: Every morning, I weigh myself. Today, I found that my scale is inaccurate, and understates my weight by 5%. Therefore, if I am to predict my weight in the future, I must account for the inaccuracy of the scale. Related Example: There exist police statistics for arrests. Studies have found that arrests are racially biased, and therefore those police statistics do not accurately represent the crime rate, only the biased arrest rate. Therefore, if I am to predict the crime rate in the future, I must account for the inaccuracy of the data set.

PeCaN 10y ago
I'm not going to get into a debate if you're going to appeal to the “actual crime rate” vs the “racially biased arrest rate”. There's not enough data to know, and too many interacting variables. It's enough for me to accept that blacks represent a disproportional amount of inmates and that black neighborhoods are frequently stuck in poverty cycles. I think the solution actually has nothing to do with correcting variables for race or anything of the sort; rather, help people—regardless of race—out of poverty and the statistics will level out. I'm extremely skeptical that arrest “bias” is skin color related. Also, citation needed for “studies have found that arrests are racially biased”. That is entirely too hard to measure. Well, it's not that I doubt such studies exist, I simply doubt they're valid in any way, and there are probably other studies with the opposite conclusion.
programmarchy 10y ago
But the problem you describe isn't biased data, it's state law enforcement agencies practicing discrimination. Seems like it'd be more effective to strike at the root, not the branches.
- MereInterest 10y ago
  Whether or not data are biased depends on what you want to measure. If you are trying to measure the arrest rate, then the arrest rates are an unbiased predictor. If you are trying to measure the true crime rate, then the arrest rates are a biased predictor. I certainly agree that racially motivated arrests are a large issue, and should be dealt with. Avoiding misinterpreting data is a widespread issue that in these cases happens to have prejudicial consequences.