Hacker News Reader

Comment by tgsovlerkhgsel | Hacker News Reader

Based on the visualization the algorithm seems to be:

If you input that you expect to be unemployed for a short/long time, it returns that assumption. Only if you say you don't know...

Immigrant/descendant: High risk of long term unemployment, right away. (Edit: Apparently except non-western immigrants, based on the source code!)

Age is the next check that can classify you as high risk, I didn't look for the threshold (Edit: 56, according to the source). As long as you're below that, the deciding factor is...

How long were you employed out of the past 36 months? 12 months is apparently good. (Edit: <2 is the threshold).

It seems like most of the questions don't seem to be considered (if the visualization is to be believed) and it's a very simple "if x return foo" style algorithm, not some fancy scorer/ML.

Look for function setValues() in https://ledighedsalgoritmen.dk/p5vis.js

View on HN

> "if x return foo" style algorithm, not some fancy scorer/ML

It's a decision tree. Decision trees falls under ML. The thresholds are learned from data.

I would imagine they had to use a decision tree in order to tell users and case workers _why_ they are at risk.

The algorithm has been heavily criticized for putting basically all non-western immigrants into the "high risc"-group, but if the job market historically hasn't been willing to hire high performing, non-western immigrants, isn't it the whole point of building an algorithm from data? To shed light on this?

Feels like a decision tree is not the kind of algorithm that can solve this problem due to its multidimensional nature. It's very vulnerable to Simpson's paradox.

It may be right most of the time if we lump everyone in the same bag, but that is not helpful at all.

Also, having the self-evaluated unemployment risk the first deciding factor seems to defeat the purpose.

A fixed decision tree isn't machine learning. If the tree can be updated or generated by code, then sure, but a human just writing a decision tree has nothing to do with machine learning.

The interesting thing is that at least the version on this website does NOT put non-Western immigrants into the high risk group. It puts Western immigrants (and descendants) there, and non-Western descendants, but not non-Western immigrants themselves!

> Immigrant/descendant: High risk of long term unemployment, right away. (Edit: Apparently except non-western immigrants, based on the source code!)

The latter part is probably on the basis that non-Western immigrants mostly needed skilled worker visas, whereas descendents and 'Westerners' (which apparently maps to "EU members" in this case) don't. Though as with most of the rest of it I'd struggle to see this as a conclusion from advanced modelling and not just a stereotype kept relevant largely by not asking the descendent or 'Westerner' what type of work they'd been doing and how long they'd worked in Denmark

There are (compared to SV) very few asians in Denmark. They also integrate well so the Danish "I don't have a job" system isn't focused on them.

It is a long-standing taboo to split "non-Western" into "middle east" and "asian". Avoiding the split makes the bad numbers less noticeable.

> Look for function setValues() in https://ledighedsalgoritmen.dk/p5vis.js

It seems you've found the algorithm right there. Not only is this algorithm the good old mega-nested-if-test-of-hard-coded-answers for a few of the questions asked. It basically checks if you think you are at risk, then puts you "atRisk" if you are older than 56/about to retire/an immigrant/pregnant.

Sadly; that is probably enough to report relatively correct values.

There is another way to look at it; a student was able to visualize what they wanted with minimal investment in learning secondary tools to their actual field.

Similar issue is often misattributed in other places like game engines aswell. For example Unity is often labeled as lower quality because of the correlation with bad code quality in many of the games made with it, while forgetting that this is precisely because it's so easy to get started with, more new developers choose it

I don't know that it's so sad. If it replaced humans putting their fingers in the air and then pulling an answer out of their ass, it's probably a strong improvement.

Reply on news.ycombinator.com