Hacker News Reader

Comment by djohnston | Hacker News Reader

[flagged]

The goal is to avoid penalizing people for their skin color, or for gender/sex/ethnicity/whatever. If some group have higher rate of welfare fraud, the fair/unbiased system must keep false positives for that group at the same level as for general population. Ideally there should be no false positives at all, because they are costly for people, who were marked wrongly, but sadly real systems are not like that. So these false positives have to be spread over all groups proportionally to sizes of the groups.

Though the situation is more complex than that. What I described is named "False Positive Share" in the article (or at least I think so), but the article discusses other metrics too.

The problem is that the policy should make the world better, but if the policy penalizes some groups for law breaking, then it can push these groups to break the law even more. It is possible to create biases this way, and it is possible to do it accidentally. Or, rather, it is hard not to do it accidentally.

I'd recommend to read "Against Prediction", it has a lot of examples how it works. For example, biased False Negatives are also bad, they make it easier for some groups to break the law.

>The goal is to avoid penalizing people for their skin color [...]

That's not correct. The goal is to identify and flag fraud cases. If one group has a higher likelihood to perform that, then this will show up in the data. The solution should not be to change the data but educate that group to change their behavior.

Please note that I have neither mentioned any specific group and do not have a specific group in mind. However, an example for such a group that I have seen in my professional life could be female 20 year old CEOs of construction companies (often connected to organized crime)

> The solution should not be to change the data but educate that group to change their behavior.

1. This is easier to say than to do.

2. In reality what you see is a correlation. If you try to educate all 20 year old females to not become a connected to organized crime CEOs of construction companies, your efforts will be wasted with 99% of these people, because they are either not connected to organized crime or are not going to become CEOs. Moreover the very your efforts will lead to a discrimination of 20 years old females, if not due to public perception of them, then because you've just increased difficulties for them of becoming a CEO.

> The goal is to identify and flag fraud cases.

Not quite. The goal is to reduce the amount of fraud cases. To identify and flag is a method of achieving that goal. But policymakers has a lot of other goals, like avoiding discrimination or reducing rate of murders. By focusing on one goal policymakers might undermine other goals.

As a side (almost methaphisical) note: it is one of the reason, why techies are bad at social problems. Their math education taught them to ignore all irrelevant details, when dealing with a problem, but society is a big complex system where everything is connected, so in general you can't ignore anything, because everything is relevant. But education have the upper hand, so techies tend to throw away as much complexity as it is needed to make the problem solvable. They will never accept that they don't know how to solve a problem.

"your efforts will lead to a discrimination of 20 years old females"

I'd think that this is an extremely far-fetched example that fails at basic logic. Just because a very specific scenario will be flagged does not mean that this scenario is generalized to all CEOs, all females, all 20 year olds.

I think their point is that a lot of us, upon hearing "group A tend to exhibit more of some negative trait X than some other group B" mentally start to associate A with X and this creates a social stigma - just because how our brains work.

I wish there'd be some way to phrase such statements in a nonjudgmental way, without introducing a perception bias...

> The goal is to reduce the amount of fraud cases.

I'm sorry but I fail to see how have you reached this conclusion. Can you please elaborate? The way I understand it, a detection system cannot affect anything about its inputs - you need a feedback loop for that to happen. And I don't see anything like that within the scope of the project as covered by the article.

In practice, investigations tend to find the results for which the investigation was started. At the beginning of the article, it was also suggested that such investigations in Amsterdam found no higher rate of actual fraud amongst the groups which were targeted more frequently via implicit bias by human reviewers.

In North America, we know that white people use hard drugs at a slightly higher rate than non-whites. However, the arrest and conviction rate of hard drug users is multiples higher for non-white people than whites. (I mention North America because similar data exist for both Canada and the USA, but the exact ratios and which groups are negatively impacted differ.)

Similarly, when it comes to accusations of welfare fraud, there is substantial bias in the investigations of non-whites and there are deep-seated racist stereotypes (thanks for that, Reagan) that don't hold up to scrutiny especially when the proportion of welfare recipients is slightly higher amongst whites than amongst non-whites[1].

So…saying that the goal is to avoid penalizing people for [innate characteristics] is more correct and a better use of time. The city of Amsterdam already knew that its fraud investigations were flawed.

[1] In the US based on 2022 data, https://www.census.gov/library/stories/2022/05/who-is-receiv... shows that excluding Medicaid/CHIP, the rate of welfare is higher for whites.

The better definition of equal performance would obviously be that the metrics for the detector - accuracy or false positive rate etc would be the same for all groups.

I won't comment on why it's defined the way that it is.

Edit: it looks like they define several metrics, including ones like I mention above that consider performance and at least one based on what number or percentage is flagged in each group.

Or that the error distributions are equal across groups. That way you could still detect that one group is committing fraud at a higher rate, but false positives/negatives occur at the same rate in each group

There are multiple different ways to measure performance. If different groups have different rates of whatever you're predicting, it is not possible to have all of the different ways of measuring performance agree on whether your model is fair or not.

> Why would you assume that all groupings of people commit welfare fraud at the same rate?

Because the goal is NOT just wiping out fraud, but, instead, minimizing harm or possibly maximizing positive results.

Minimizing fraud is super easy--just don't give out any benefits. No fraud--problem solved.

That's not the final goal, though. As such, the ideal amount of fraud is somewhere above zero. We want to avoid falsely penalizing people who, practically be definition, probably don't have the resources to fight the false classification. And we want to minimize the amount of aid resources we use policing said aid.

The goal is to find a balance. Is helping 100 people but carrying 1 fraudster a good tradeoff? Should it be 1000? Should it be 10? Well, that's a political discussion.

Oh no you really stepped in it now

Yes it is. This is some ideal world thinking, that has nothing to do with reality and is easily falsifiable, but only if you want to see the real world.

> Why would you assume that all groupings of people commit welfare fraud at the same rate?

What's the alternative? It's an unattainable statistic, the people who get away with crime. Instead, what ends up getting used is the fraud rates under the old system, or ad hoc rules of thumb based in bigoted anecdotes.

So instead you delcare that you don't think that ethnicity is in and of itself a cause of fraud. Even if there may be any number of characteristics that tend to indicate or motivate fraud that are seen more in one specified ethnicity than another (poverty, etc.), and even though we should expect that to lead to more fraud. We can choose to say that those characteristics lead to fraud, rather than the ethnicity, and put that out of scope.

Then we can say that this algorithm isn't meant to solve multiculturalism, it's meant hopefully not to exacerbate the problems with it. If one wants to get rid of weird immigrants, non-whites, or non-Christians, just do it, instead of automating a system to be bigoted.

Also, going after the marginal increase of rates of fraud through defining groups that represent a small portion of the whole is likely to be a waste of money. If 90% of people commit fraud at a 5% rate and 10% commit it at a 10% rate, where should you be spending your time?

One of these days, I’m still hopeful, we will figure out that behaviors are taught, usually by parents. Intentionally or accidentally, kids learn from what they see.

I don’t care what nationality you are, or what your skin color happens to be, the root cause is how kids are reared.

In my head it’s so simple.

Not just kids - it's about one's whole life, including the adulthood, to the very last moments. We tend to change a lot over the courses of our lives, constantly being affected by our surroundings.

Using the word “retarded” like that negates any other point you tried to make.

Reply on news.ycombinator.com