Hacker News Reader

Comment by msellout | Hacker News Reader

Why not use logistic regression to estimate likelihood of signup rather than clustering the population into 5 groups? I'd imagine the resulting predictive model would not only have more direct methods for measuring prediction error, but also provide more business insight.

View on HN

Absent underlying qualitative differences, there is rarely a good reason to break a continuous distribution into discrete groups for modeling purposes or for model performance evaluation[1].

But for human consumption (and I believe this article is an example of that), it can help. It's basically an ocular Hosmer-Lemeshow; not a rigorous or even consistent approach to model performance evaluation, but often interesting to those consuming the model's output. For example, we do it here to give students a sense of what their chances have meant historically: https://www.parchment.com/c/college/college-1404-University-...

[1] See the Hosmer-Lemeshow test, now uncommonly used.

Oh, I'm a big fan of logistic regressions. And this model actually has some logistic regression in the mix. But I just used this break down of 5 groups to easily report the performance of the model. I can also report a gini but it is way more visceral to see "this is who I predicted to be my top performing group, this is how big that group was, and this is how they actually performed."

In general your work may be complicated/sophisticated. But your results need to be simple.

Note that probabilities estimates from logistic regression are often not well calibrated. This typically doesn't matter if you're using logistic regression to create a classifier. However, if you are making cost-sensitive decisions based on probabilities predicted by logistic regression you make run into trouble. There are a number of publications on calibrating probabilities that suggest fixes.

All that said, I doubt it makes a difference in this case.

Reply on news.ycombinator.com