Hacker News Reader

This would be much more interesting if the features weren't anonymized.

At this point this is just a widest/deepest neural net competition on some unknown bunch of features.

100% this. Anonymised features remove any hint of lateral thinking that is the essence of science, and just focuses on the pure numbers. This is dangerous. One of my data scientists once ran up a $2000 bill on training a CNN on what was essentially a speed = distance/time calculation, and was really proud when he got 92% accuracy. A real facepalm moment.

I was just about to post the same thing. I'm a statistician in my day job and the raw math is only one part of building a model. Human judgement (while often flawed) can be key to improving actual model performance.

Reply on news.ycombinator.com