Human level Performance?

I was going through some courses in coursera, which was talking about human level performance, which makes sense BUT what about a task like predicting house prices?

Would that that still be close to Bays Error?
What would be human level performance in this case?

Hi @darthShana
Welcome to the community!

Human level performance usually used is the cases of the pixelated images that may the human couldn’t see the image and classify what it is and comparing the highest Accuracy of the model you built to the highest value of human level performance, but in the prediction projects I think that the human level performance would be so good(above 98%), so that the human level performance is used much in the image classification or image recognition

Cheers,
Abdelrahman

1 Like

The definition of Bayes Error is that it is the least possible error that a system could theoretically have on a given problem. It’s often not easy to prove what that value is, but you can use Human Error as a proxy for that. We know that Bayes Error <= Human Error, by definition. Of course we also have no way to know how big the gap is there, but frequently that’s the best we can do.

In the case of housing prices, you can find plenty of predictions from human real estate agents. You can argue that the listing price of any house that goes on the market is the selling agent’s best guess as to what it should sell for. Of course there are games that they play sometimes based on how they think the market will react to underpricing a listing in a hot market. You could then use the difference between the actual sale price and the listing price as the basis for calculating the human error in this case. There are tons of data like that available on any realtor’s website.

1 Like