3 ms·
Because you hose people who are not actually high-risk but, for whatever the reason, chose to hang out with people who are high-risk. This is not really a socia
by sliverstorm 13y ago
Because you hose people who are not actually high-risk but, for whatever the reason, chose to hang out with people who are high-risk. This is not really a socially acceptable thing to do, because you are effectively punishing those individuals for their choice in friends. While the First Amendment may not apply here, I'd say most people value the spirit of "freedom of association" and thus I expect it to have been applied elsewhere.
- markdown 13y agoI wonder how Obama would have been classified in his "Community Organizer" days. if job == 'Lawyer': score + 100 if poor_people_friends > 150: score - 100
- Houshalter 13y agoThat's not how machine learning. I mean no one just programs "if poor_people_friends" or keeps a running score. They just feed the machine the data.
- markdown 13y agoAnd the machine just guesses a score? Are you saying the machine doesn't actually rank you according to the data you give it?
- Houshalter 13y agoI wouldn't equate not giving someone a loan because they are at high risk as "punishment". But anyways it's just a data point. Why does it matter what the data point is, if it helps make more accurate predictions? No it's not socially acceptable to use certain data points to make decisions, but this has good justification behind it. People will irrationally weigh stuff like race or gender to make decisions even when it doesn't matter. But this is being done by a completely unbiased computer program. And they wouldn't use it if it made their predictions worse. So that reasoning of the predictor being irrational doesn't apply in this case.
- sliverstorm 13y agothis is being done by a completely unbiased computer program And I'm sure the program was written by a completely unbiased computer, too!
- Houshalter 13y agoI highly doubt they would intentionally make their models more inaccurate to reflect their own biases. Machine learning algorithms are pretty unbiased as far as weighing the correlation of specific features goes.
- ekianjo 13y agoProbably you have never seen Bayesian modeling in action. It doesn't care what data you throw at it, it will establish models out of it. The theory IS neutral.