Hacker News Reader

There's also bias in the data itself. That's the difficult thing to avoid. Even down to how we phrase a question, who we collect the data from, it all introduces a bias unless we're literally harvesting all data from every human being and using that for our models. There's no way to get rid of the bias, even if we take out the nudges.

View on HN

How about you select a representative (i.e. random and statistically significant) sample of the population and then ask them their opinions about certain (especially controversial) parts of your data, and then weigh your data according to these opinions?

That's just an idea that occurred to me (in 30 seconds of thought) which could probably make the training data significantly more unbiased.

But I'm sure there are research scientists who can come up with better methods for sampling data in a more unbiased fashion.

Note that this is not an all or nothing approach. Your training data could presumably be 100% biased or 0% biased, but also any value in-between.

The goal is to try to make it as close to 0% biased as feasible, given whatever effort you're comfortable expending.

I'm not sure why you are conflating bias with the "the statistical likelihood that a belief is held by someone". They have nothing to do with each other.

Reply on news.ycombinator.com