Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
baogorek
left a comment
There was a problem hiding this comment.
Hi Maria, I just ran the test and it ran very smoothly and I saw the sparse output. I see that you were able to leverage an open-source parameter optimizer (Optuna), which was smart. One thing I would like to see is a bit of information about the holdouts once tuning is one. There should be an error for each of the holdout sets and a way to see which targets were held out. I'm probably not going to review every line of code here, but if I could see that, I would feel very good about putting my checkmark here.
baogorek
left a comment
There was a problem hiding this comment.
There's a lot in here, and I don't want to block useful functionality from going in. I am worried about data leakage from the training folds into the holdout folds, which is virtually guaranteed to happen given the way we define targets. All I would ask is that there be some warning added so the user can see it and we will be reminded to make it more robust in the future.
Fix #77
Fix #76
This enables tuning the L0 parameters to achieve a combination of loss reduction, estimates being within 10% of targets, and sparsity (proportion of weights that are 0). The objective function combines the three, and the parameter "objectives_contribution" allows changing how much we want to prioritize each objective, making it flexible to tune for different needs.
It also supports target holdout.
n_holdout_setscan be set to 1 with aholdout_fractionof 0 if we dont want to holdout targets.