Expt/regularization of parameters #157
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Is |
Yep. It's the number of reviews used for training. |
|
Alright. I'd like to see you benchmark this on 1000-2000 collections with 3-5 different gammas. |
|
|
So, no significant impact? Weird. Do a few more with even larger gammas. |
|
The tendency shows that large gamma makes the model worse. |
|
Hm. How about trying it only on collections with <1000 reviews? |
|
One more thing: aside from the weighted mean, add the 99th percentile for logloss and RMSE, like this: I want to see how much regularization helps in the worst cases. |
The LogLoss is improved but the RMSE(bins) is worse on average. Here is the result on all collections: In my opinion, it's OK to merge it because it improves log loss and AUC at least without increasing RMSE(bins). |
|
Please try more gammas first. |
|
What's the purpose of this? Wouldn't regularization just prevent users with small decks (presets) from generating the parameters that are most optimal for them? |
|
Perhaps for users with small collections staying close to default parameters is actually more optimal due to high amounts of random noise isn such colelctions. |
|
One disadvantage is that if I split a deck into two (clearly knowing that they would benefit from having different parameters), the new decks would be less likely to show that improvement in the RMSE because the smaller deck size will decrease the amount of deviation that is permitted from the default parameters. |
For this particular change, I think that it would be better to first split the database into small and large collections (based on some arbitrary cutoff) and then perform statistical significance tests on both groups separately. |
In small sample size, it's high likely to overfit the data, which weaken the model's ability to generalize. Regularization could increase generalization in this case. |
In summary:
|
|
Try gammas even higher than 2.0 |
|
|
Alright, let's use gamma=2 then |
Expertium
approved these changes
Jan 9, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.