Attacking discrimination with smarter machine learning • Hacker News Board

Re: Attacking discrimination with smarter machine learning

#141

Post by qb45 » Mon, Nov 21, 2016, 6:27 PM UTC

idanoeman wrote: post #140

tptacek wrote: post #116

Earlier quoted context omitted.

Generally we're most of us in society pretty comfortable with paying for the continuation of the species. There are exceptions to every rule, of course; some of us (not including me) are unhappy we pay to pave the streets, too.

I don't think that we need to be subsidizing babies when the world population is 7.4 billion.

The problem is that retirement is a Ponzi scheme and it's not clear if the current 3rd world children will care (or even have the means to care) about our 1st world problems in the future.

Re: Attacking discrimination with smarter machine learning

#142

Post by yummyfajitas » Mon, Nov 21, 2016, 6:32 PM UTC

tptacek wrote: post #24

mvindahl wrote: post #22

Well. At the end of the day, the companies will pick thresholds and rates to maximize their profits, based upon the data that they have available. They don't know every detail of our personal lives -- which would also be kind of unsettling -- so they have to resort to a simplified picture. Simplifications are always prejudiced at the individual level but if the prejudice is reflected in the numbers, it's balanced. It…

The difference with machine learning is that the model isn't designed by humans, through an actuarial process we can keep our brains wrapped around. It's a black box. We are OK attributing "crash risk" to young male drivers, because we can observe both that they are as a cohort statistically likely to crash and also understand why that would be the case. On the other hand, we're not comfortable with the idea that a c…

No, bogus correlations and inaccurate conclusions are explicitly NOT the computer science problem being worked on here. You are trying to frame the problem as algorithms wrongly classifying people. That's just a standard statistics problem (called "overfitting") and has nothing to do with fairness.

The problem being worked on here is "what if Armenians shouldn't get car loans because they don't pay them back as much as other groups?" I.e., algorithms rightly classifying people leads to results that we believe are "unfair".

This research is illustrating that you can't simultaneously have accuracy and fairness. You need to explicitly decide how much accuracy you are willing to give up to get fairness. I.e. it's computing the tradeoffs needed to evaluate the ethical question: how many Armenian deadbeats should you extend credit to in order to be "fair"?

Go play with the simulation to see. The various fairness criteria all achieve lower than maximal profits.

Re: Attacking discrimination with smarter machine learning

#143

Post by dang » Mon, Nov 21, 2016, 6:34 PM UTC

chronic6l wrote: post #113

Blacks (or African Americans) are less intelligent than other ethnicities, in average. We have information that satisfies your two criteria of supporting class information. Yet people on HN (liberals/progressives) will get upset with this fact. Supporting classes: 1. Through employment, high school graduation, incarceration, and homicide rates. 2. Due to black cultural, male macho-independence, discrimination from no…

Racewar threads are off topic on the grounds of ultimate tediousness.

All it leads to is flamewars, and no flamewar will ever resolve any of it. Meanwhile it poisons what we do care about: intellectual curiosity, and civil, substantive discussion.

Gwern has a nice bit somewhere about how you can either devote your entire life to this stuff or give up. Most of us, the vast majority, have given up. The rest of you true believers can take your stone tablets with "Science" engraved on them to some place where the audience actually likes it.

We detached this subthread from https://news.ycombinator.com/item?id=13005441 and marked it off-topic.

(Edit: your comment history has blatantly been breaking the HN guidelines about other things as well. We ban accounts that do that, so please stop doing that.)

Re: Attacking discrimination with smarter machine learning

#144

Post by yummyfajitas » Mon, Nov 21, 2016, 6:51 PM UTC

joncrocks wrote: post #109

losvedir wrote: post #64

Earlier quoted context omitted.

I don't get the impression that most people are afraid of irrelevant correlations, but rather relevant ones that nevertheless disproportionately affect certain populations. To use your Armenian example, it could be that while being Armenian doesn't actually affect your driving, a "true" model could still end up being bad for Armenians if being Armenian is correlated with the things that actually do affect crash risk.…

I think that people are afraid of conditions that one can't control, and are correlated with negative characteristics. Lets say that I am male, and am young. Lets say that as a young male driver, I am likely to be a cause of an accident 20% of the time i.e. 1 in 5 young male drivers will cause an accident. Lets say as a young female driver, I am likely to be the cause of an accident 5% of the time. So on the face of…

No, that's not what this research is about at all.

This research is about what happens when combining the recklessness score with gender-based distributions provides more accurate results than recklessness alone. Many people consider this to be "unfair". This article shows how much profit you need to sacrifice to get fairness.

Re: Attacking discrimination with smarter machine learning

#145

Post by PeterisP » Mon, Nov 21, 2016, 6:55 PM UTC

amelius wrote: post #4

But how can we find out if a company uses ML in a non-discriminating way? If we cannot see it, and check it, then there is no incentive for companies to use it. My guess is that at most companies spending time on making an algorithm non-discriminating will be viewed as a waste of time and money.

Since the original article shows nice examples that there is no single "non-discriminatory way" possible - at the very least, you have to do the classic tradeoff between equal opportunity (and discriminatory outcomes) or equal outcomes (and discriminatory opportunities), you can safely assume that all companies use ML in somewhat discriminating way. At the very least, if they ask your ZIP code and it "matters" in their decision system in any way, then that's a positive sign of discrimination (one way or another) since it's so correlated with race among other things.

Re: Attacking discrimination with smarter machine learning

#146

Post by dang » Mon, Nov 21, 2016, 6:56 PM UTC

powertower wrote: post #137

tptacek wrote: post #127

Earlier quoted context omitted.

No, he doesn't. He has a mixture of early Jensenist psychometric research that has been superseded, and neo-phrenologists like Rushton that have been discredited. But of course it's easy to drop little bombs like this into threads and put the onus on other people to explain the science. That's what makes it trolling. Whatever snappy response you have for this, please spare us. There is a reason "the subject of racial…

We do know that heredity/genetics plays a major role in your physical structure - giving you everything from having two arms and two legs, to making you susceptible to some types of cancer. And there have been many many studies, using millions of subjects, twins, various racial groups, etc, that show the same correlation on your non-physical aptitudes (and their ranges). The fact is if you sample the different racial…

HN is not a place for political and ideological battle. We ban accounts that primarily use it this way, so please stop using it this way.

Re: Attacking discrimination with smarter machine learning

#147

Post by yummyfajitas » Mon, Nov 21, 2016, 7:02 PM UTC

mspeedy wrote: post #90

tomp wrote: post #49

Earlier quoted context omitted.

I think it's useful to figure out why there are discrepancies between two groups. For example, let's take blacks in the US. The data tells you that a black person is more likely to be a criminal than a white person. There are two possible reasons for this: (1) blacks are more prone to crime, or (2) blacks are more likely to live in circumstances that make them criminals. With access to only anecdotal data, I strongly…

> if you took into account enough circumstances (e.g. single parent, school district, income level, parents' wealth) you'd be able to remove race from your model and still arrive to the "equal opportunity" result. The problem is, when given access to a large number of classifiers, some of which have inevitably been affected by a pre-existing racial bias, a black box machine learning algorithm will likely become discr…

The problem is, when given access to a large number of classifiers, some of which have inevitably been affected by a pre-existing racial bias, a black box machine learning algorithm will likely become discriminatory as well if race is not in some way represented and equalized.

This is simply not true. Black box machine learning algorithms will have the tendency to correct bias in their inputs. Insofar as they do systematically deliver wrong answers, this is actually called "variance" and has no particular sign. It's just as likely to be biased in favor of $protected_class as against that class.

https://www.chrisstucchio.com/blog/2016/alien_intelligences_...

Also, you do know that Pro Publica's R script actually found no bias, right? The bias was actually in the selection of anecdotes in their article, which obscured the fact that their statistical analysis could not reject the null hypothesis.

https://www.chrisstucchio.com/blog/2016/propublica_is_lying....

Re: Attacking discrimination with smarter machine learning

#148

Post by dang » Mon, Nov 21, 2016, 7:11 PM UTC

alvarosm wrote: post #60

alvarosm wrote: post #28

This is how it should be: "Max Profit. The most profitable, since there are no constraints. But the two groups have different thresholds, meaning they are held to different standards." Here's the big fallacy: "the two groups have different thresholds, meaning they are held to different standards." They are not held to different standards because they're different groups, but because of other reasons that indicate dif…

Any SJW care to explain why I'm wrong instead of downvoting? thanks! this is a lot of fun :D

Since you've ignored our repeated requests to stop breaking the HN guidelines, we've banned your account.

If you don't want to be banned, you're welcome to email hn@ycombinator.com. We're happy to unban people if they give us reason to believe they'll only post civil, substantive comments in the future.

Re: Attacking discrimination with smarter machine learning

#149

Post by yummyfajitas » Mon, Nov 21, 2016, 7:12 PM UTC

riskable wrote: post #11

amelius wrote: post #4

But how can we find out if a company uses ML in a non-discriminating way? If we cannot see it, and check it, then there is no incentive for companies to use it. My guess is that at most companies spending time on making an algorithm non-discriminating will be viewed as a waste of time and money.

I work for a big bank... As long as people freak out when banks are found to be discriminating they will do their best to not discriminate. Banks are built on trustworthiness. Having your bank's name in the headlines for discriminatory practices can have a severe negative impact on trustworthiness. They have a whole teams of people devoted to this topic and every year at most banks every employee has to learn about,…

The problem here is that non-discrimination costs you money (in practice: lots!), because it requires you to issue bad loans to non-Asian minorities.

Secondly, it isn't mathematically possible to be "nondiscriminatory" - there are multiple definitions of that term and they are mutually conflicting. For example, as this article shows, "equal outcomes", "equal opportunity" and "equal treatment" (group unaware) don't make the same decisions.

So no matter which definition of "fair" you choose, some intrepid reporter can choose a different definition and then write a clickbait article calling you racist.

Re: Attacking discrimination with smarter machine learning

#150

Post by yummyfajitas » Mon, Nov 21, 2016, 7:18 PM UTC

d--b wrote: post #89

alvarosm wrote: post #46

You realize you're advocating totalitarianism, right?

Race and ability to repay a loan, are - and this is scientifically proven - uncorrelated. What is correlated is your social status and your ability to repay a loan. However, there is a problem because the population of the US is not uniform. Because of history, some races were less wealthy than others. So far, this makes sense I suppose. Now, talking about machine learning. The big specificity of machine learning (wh…

Can you cite the scientific proof? From what I've seen, race is correlated with default probability after taking other obvious factors into account (e.g. income, education, single motherhood).

Figure 7 of the underlying paper we are discussing shows exactly that - a white/asian person with a 700 FICO score has about a 10% chance of defaulting, while a black person with a 700 FICO has about a 15% chance of default. (Those numbers are eyeballed from the graph, might not be that accurate.)

I can't cite all the information I have on this, but I've found a blog post which analyzes similar data and derives similar relationships in publicly available data:

https://randomcriticalanalysis.wordpress.com/2015/11/22/on-t...