The sedan vs. pickup truck example (you read the article right?) didn't sound like a distinction of material wealth to me. It seems more related to what kinds of attitudes/values/views/etc. that kind of person has. e.g. the obvious (naive?) possibility here is that Democrats don't like gas guzzlers due to a bigger concern about the environment.
I would be very curious if you'd see these distinctions even in neighborhoods where occupational reasons don't exist. My hunch is that you still would, but I don't know/haven't looked into it.
Am I the only one who is a little offended that just because you own a pickup truck then you must be conservative, and if you own a Tesla you must be liberal? Seems to me that is more stereotyping than lot's of things that social justice warriors get outraged over.
> just because you own a pickup truck then you must be conservative
Could you point us to whatever sentence in the article that made or even suggested this? I didn't see any part of the article stating that everyone who has a pickup truck is conservative.
> For instance, if the number of sedans in a neighborhood is greater than the number of pickups, there is an 88 percent chance that the precinct will vote Democratic. Transpose those numbers to have more pickups than sedans and there is an 82 percent chance a precinct will vote Republican.
Or that if you feed data which contains clear indicators of bias/prejudice into your ML system, it will discover those biases and prejudices and learn and incorporate them.
Got a training set which dates to when banks redlined black people? Congratulations, the model just learned not to make loans to black people! And now it's "objective", because no humans are involved in making its decisions!
If you define an objective function based on some set of predictors, none of which contains a Black variable then it is not possible for your model to discriminate against Black people because it doesn’t know if anyone is black.
Assuming, ridiculously conservatively that redlining ended in 1979 any dataset will have 37 non redlined years in it, 1980-2017. Given that the further away in time a datapoint is the less weight it’s given in predicting anything redlining will have bigger all impact on any model.
Asian people in the US on average have higher credit scores than white people. This is not anti-white discrimination because white and Asian are not used to determine credit score, it’s based on on-time payment, capacity used, length of credit history, types of credit used and past credit applications.
There are many wonderful MOOCs and textbooks on ML, AI and statistics. If you want to understand how statistical modelling works I can recommend the Georgia TechX MicroMasters in Analytics.
Don’t listen to statistically illiterate journalists.
Redlining is only one of the many racists housing practices that have existed and continue to exist today.l
> in 1979
Have you even done any research on this topic? From a famous article[1] about the history of racism in housing:
>> In 2011, Bank of America agreed to pay $355 million to settle charges of discrimination against its Countrywide unit. The following year, Wells Fargo settled its discrimination suit for more than $175 million. But the damage had been done. In 2009, half the properties in Baltimore whose owners had been granted loans by Wells Fargo between 2005 and 2008 were vacant; 71 percent of these properties were in predominantly black neighborhoods.
> 37 non redlined years
Those are also years of sub-standard schools, poor job prospects, and (in some cities) toxic levels of lead paint. Growing up in that kind of environment can limit your options severely.
> it is not possible for your model to discriminate against Black people because it doesn’t know if anyone is black.
I always like to assume ignorance over malice, so I will simply suggest that you should do a lot more research before making this kind of claim.
ML algorithms are only going to have evidence-based biases (although I guess with a motivated choice of training data you can make the evidence have whatever bias you want)
That includes historic institutional biases. For example, many US cities are still highly segregated by race. This was done intentionally as a workaround for forced-integration rulings with indirect methods such as redlining, blockbusting, and restrictive covenants. While most of these methods were eventually banned, the results of those methods still exist today[1].
Since school funding usually depends on property taxes, these policies also impacted the school system. It may be a fact that one group of people performs worse on some metric, but how much of that was caused by racist housing practices decades earlier? Unless you're accounting for all of the embedded biases already present in the facts, you are laundering those biases into your results.
Classifying anything related to people is a political act; it isn't possible to be apolitical and "just make a classifier". The classifier either accounts for the biases already present in the data, or they are present in the output, obscured (laundered) behind an opaque wall of machine learning.
Unjust policies laundered those biases into reality, at which point they stop being biases and start being facts with unjust causes.
Whitewashing the training data to make predictions based on a hypothetical injustice-free world where redlining never happened will just give you garbage predictions.
I agree that it's hard to get good predictions. The point, however, is that ignoring systemic bias and other historic influences is worse.
> and start being facts with unjust causes.
Of course. That means you know the data contains an "unjust" bias. Do you want to account for that bias, or do you want to proceed anyway risking problems (e.g. Simpson's Paradox).
There are more confounding factors in heaven and earth, closeparen, than are dreamt of in your model.
Confounding factors undermine causal arguments. You can build a perfectly useful model without knowing about the hidden variables.
A risk-management approach to the criminal justice system doesn’t call for fudge factors. It’s fundamentally unacceptable. Risk management is all about using people’s demographics against them (see: all insurance pricing). Wokeness on the part of the risk managers doesn’t fix that. The criminal justice system needs to treat people as individuals and not classes - this requires that it steer clear of evidence-based risk assessment.
That’s a description of reality. 88% of neighborhoods that had more sedans than pickups lean blue. I don’t see why that would be offensive? Seems in line with each group’s general culture.
You can't ban any statistical reasoning about humans because it's sometimes wrong.
"We stopped testing the medicine because even though it cured 99% of cases those that weren't cured were offended by the possibility that someone might assume that it would cure them"
Statistical reasoning becomes prejudice when you act on it in a way that is harmful to society or you overgeneralize it. If you try to ban it then you're just avoiding emotionally uncomfortable truths.
It's similar to how the distribution of height in men and women is different, ie, women tend to be smaller (IIRC), it has no bearing on the expression of the measures values on an individual.
There is a probability that you are conservative if you have a pickup truck but this doesn't mean you are conservative if you own a pickup truck.
And that’s, girls and boys, how a few PhDs back in the 10’s, either by stupidity or lack of ethics, laid the foundation for the big brother state we have today.
How do they ascertain that the cars belong to the suburb? Would a suburb be ranked differently if for a month there were civil works that meant a lot of contractor vehicles etc were parked up?
What about if the local Ferrari club decides to go for a cruise and all meet at a coffee shop in the neighbouring suburb every week? What if the suburb was located in a cold region, where it's more practical for most people to drive a SUV in winter?
What about the republicans who drive Civics, and the republicans who drive Audis? What about the democrats that drive Toyota Tacomas, or the democrats who drive Volts?
I mean, what would it make of me? I have a toyota hilux for work, a mx5 (miata) for track days, and a subaru wrx sti for driving to the shops. I wonder what my political leaning would be considered to be?
Certainly. My main concern is the line: "on par with some information that takes billions of dollars to obtain via census surveys"...
As soon as you have a researcher make this sort of claim, a politician will inevitably see it as "Don't spend money on census research, do this instead" ...
I wouldn't put it past a group of Fezza owners to be conspiring about something at any given time :P
But to be serious, my point was more about the way that they are collecting their data. If you were focusing on buildings and structures in a suburb to rate demographics, I would think that you would have more accurate results. Cars don't stay still though, and just because it's in a location at a particular time doesn't necessarily mean it lives there. The ferrari club going for coffee is just an example of affluent (types of) cars possibly throwing out results.
It's doesn't mean anything to just point at the source of an error though. You need to show that the error results in a bias (edit: or variance) on a large scale. My point here was that it seems pretty plausible for the law of large numbers to take care of little variances like this, so if you want to claim their collection was problematic, explain which errors could occur consistently so as to reach a wrong conclusion.
Those are edge cases. Nothing says that this will be accurate for every individual block 100% of the time, but for most neighborhoods, the correlation between car type and political leaning appears strong.
Not that statistical analysis like this isn't interesting, but when you have two choices - Red team or Blue team - which are also highly clustered and a ton of variables to use to correlate - pickup, sedan, hybrid, minivan - of course its going to be pretty easy to build a model that will pick the correct team with high accuracy.
I wonder if they've got access to historical streetview images, to test how stable the claimed relationship is over time.
After all, when oil prices rise you'd expect people to choose more efficient cars - not necessarily to change their voting patterns. I can't see any mention in the original paper of how they calibrated for that.
I like the change over time part. I think in the last election the "lesser of two evils" dominated, and many people who would have liked to vote Democrat but deeply didn't like Hillary had to make a horrible choice. Two states "flipped" from "blue" to "red", didn't they? Those were "rust-belt" states, so I'm very sure that there wasn't change in cars.
I think that was using a cliche phrase that has come to mean the opposite of the literal meaning. Who is wealthy enough to own a Bentley? The 1%? The 0.1%? Not me.
So what should my next car be in order to be useless to this algorithm. I don't like being categorized and sifted. Once that information is available, it is always abused. Humans never don't abuse the power they get, and I hate it. I hate being on the receiving side. It may be small abuses, but within a civilization those add up, and we are on the edge of dying a death of a thousand cuts.
I'd like to see it take into account parking habits. In my experience, it seems to be much less common to find a vehicle backed into a parking spot in more liberal areas. I wonder if this observation is correct.
This is not surprising at all. Another interesting problem that you can solve is using your social media friends' cars to predict an individual 's political leanings.
57 comments
[ 3.4 ms ] story [ 114 ms ] threadE.g. If you work in construction, or you hunt, you'd be a lot more likely to choose a pickup truck for purely practical reasons.
I have to wonder sometimes if owning a certain kind of car "just because you like it" is as reasonable a reason for ownership as I consider it to be.
Could you point us to whatever sentence in the article that made or even suggested this? I didn't see any part of the article stating that everyone who has a pickup truck is conservative.
"Everyone with a pickup truck is conservative/Republican."
from the statement
"If P(sedans) > P(pickups) then P(precinct voting Democrat) >= 88%."
?
You just learned what it means.
Got a training set which dates to when banks redlined black people? Congratulations, the model just learned not to make loans to black people! And now it's "objective", because no humans are involved in making its decisions!
Assuming, ridiculously conservatively that redlining ended in 1979 any dataset will have 37 non redlined years in it, 1980-2017. Given that the further away in time a datapoint is the less weight it’s given in predicting anything redlining will have bigger all impact on any model.
Asian people in the US on average have higher credit scores than white people. This is not anti-white discrimination because white and Asian are not used to determine credit score, it’s based on on-time payment, capacity used, length of credit history, types of credit used and past credit applications.
There are many wonderful MOOCs and textbooks on ML, AI and statistics. If you want to understand how statistical modelling works I can recommend the Georgia TechX MicroMasters in Analytics.
Don’t listen to statistically illiterate journalists.
Redlining is only one of the many racists housing practices that have existed and continue to exist today.l
> in 1979
Have you even done any research on this topic? From a famous article[1] about the history of racism in housing:
>> In 2011, Bank of America agreed to pay $355 million to settle charges of discrimination against its Countrywide unit. The following year, Wells Fargo settled its discrimination suit for more than $175 million. But the damage had been done. In 2009, half the properties in Baltimore whose owners had been granted loans by Wells Fargo between 2005 and 2008 were vacant; 71 percent of these properties were in predominantly black neighborhoods.
> 37 non redlined years
Those are also years of sub-standard schools, poor job prospects, and (in some cities) toxic levels of lead paint. Growing up in that kind of environment can limit your options severely.
> it is not possible for your model to discriminate against Black people because it doesn’t know if anyone is black.
I always like to assume ignorance over malice, so I will simply suggest that you should do a lot more research before making this kind of claim.
[1] https://www.theatlantic.com/magazine/archive/2014/06/the-cas...
Hadn't heard of it. Sounds catchy though. Could you explain the similarity? (Along with the definition of "bias" you're using?)
Intent is not required.
> only going to have evidence-based biases
That includes historic institutional biases. For example, many US cities are still highly segregated by race. This was done intentionally as a workaround for forced-integration rulings with indirect methods such as redlining, blockbusting, and restrictive covenants. While most of these methods were eventually banned, the results of those methods still exist today[1].
Since school funding usually depends on property taxes, these policies also impacted the school system. It may be a fact that one group of people performs worse on some metric, but how much of that was caused by racist housing practices decades earlier? Unless you're accounting for all of the embedded biases already present in the facts, you are laundering those biases into your results.
[1] https://www.huffingtonpost.com/entry/the-9-most-segregated-c...
Whitewashing the training data to make predictions based on a hypothetical injustice-free world where redlining never happened will just give you garbage predictions.
> and start being facts with unjust causes.
Of course. That means you know the data contains an "unjust" bias. Do you want to account for that bias, or do you want to proceed anyway risking problems (e.g. Simpson's Paradox).
There are more confounding factors in heaven and earth, closeparen, than are dreamt of in your model.
https://mathbabe.org/2017/01/04/recidivism-risk-algorithms-a...
https://mathbabe.org/2013/01/29/bill-gates-is-naive-data-is-...
A risk-management approach to the criminal justice system doesn’t call for fudge factors. It’s fundamentally unacceptable. Risk management is all about using people’s demographics against them (see: all insurance pricing). Wokeness on the part of the risk managers doesn’t fix that. The criminal justice system needs to treat people as individuals and not classes - this requires that it steer clear of evidence-based risk assessment.
Thus type of.machine learning is externalized, automated prejudice.
See also the gay detector.
"We stopped testing the medicine because even though it cured 99% of cases those that weren't cured were offended by the possibility that someone might assume that it would cure them"
Statistical reasoning becomes prejudice when you act on it in a way that is harmful to society or you overgeneralize it. If you try to ban it then you're just avoiding emotionally uncomfortable truths.
Succinct
It's similar to how the distribution of height in men and women is different, ie, women tend to be smaller (IIRC), it has no bearing on the expression of the measures values on an individual.
There is a probability that you are conservative if you have a pickup truck but this doesn't mean you are conservative if you own a pickup truck.
How do they ascertain that the cars belong to the suburb? Would a suburb be ranked differently if for a month there were civil works that meant a lot of contractor vehicles etc were parked up? What about if the local Ferrari club decides to go for a cruise and all meet at a coffee shop in the neighbouring suburb every week? What if the suburb was located in a cold region, where it's more practical for most people to drive a SUV in winter?
What about the republicans who drive Civics, and the republicans who drive Audis? What about the democrats that drive Toyota Tacomas, or the democrats who drive Volts?
I mean, what would it make of me? I have a toyota hilux for work, a mx5 (miata) for track days, and a subaru wrx sti for driving to the shops. I wonder what my political leaning would be considered to be?
All machine learning involves a cost function. How close do the results have to be to make a useful prediction in the context.
As soon as you have a researcher make this sort of claim, a politician will inevitably see it as "Don't spend money on census research, do this instead" ...
I presume they looked at more than one neighborhood? Or are you suggesting Ferrari owners have a conspiracy to do this in general?
But to be serious, my point was more about the way that they are collecting their data. If you were focusing on buildings and structures in a suburb to rate demographics, I would think that you would have more accurate results. Cars don't stay still though, and just because it's in a location at a particular time doesn't necessarily mean it lives there. The ferrari club going for coffee is just an example of affluent (types of) cars possibly throwing out results.
Maybe their conspiring is the reason they got their Ferraris in the first place ;-)
Of course you can't assert things on a single individual based on his car. Just give a probability based on all the evidence you have.
But on a great number of people (a neighborhood) it doesn't surprise me that you get pretty good results.
http://www.pnas.org/content/early/2017/11/27/1700035114.full...
even the existence of this correlation is not obvious
After all, when oil prices rise you'd expect people to choose more efficient cars - not necessarily to change their voting patterns. I can't see any mention in the original paper of how they calibrated for that.
I'm sorry, what? A Bentley is "understated"?