[Doc] max_tree_depth LGBM about arfs HOT 5 CLOSED

Hi @ThomasBury,
I agree with leaving default parameters, except what you mentioned. Actually class_weight could be skipped by setting is_unbalance=True in the LGBMClassifier

from arfs.

Hi @partiallycomplex

Thank you for your kind words.

It is indeed a good idea to limit the depth, to avoid using "specific" models in the selection process. The idea is to avoid overfitting and keep the default parameters of lightGBM as is. The authors said that the default values guarantee sane but good behaviour in most cases.

For instance, the depth is not restricted (-1) but the max number of leaves is set to 31. In most cases, it avoids overfitting. The number of estimators is by default set to 100. You could decrease it a bit if you think it leads to overfitting.

The only parameters I'd change are:

• the loss (according to the task)
• the sample_weight (if any)
• the class_weight if dealing with unbalance classification.

You can also compare the results to GrootCV which is more robust (cross-validated).
I'll try to add this in the doc, which is now live on ReadTheDocs btw ^^

from arfs.

Hi @ThomasBury,
I agree with leaving default parameters, except what you mentioned. Actually class_weight could be skipped by setting is_unbalance=True in the LGBMClassifier

Yes, both can be used for unbalanced classification. Whatever works best for your project 🙂

from arfs.

Hi @ThomasBury, class_weight is needed for multi-class classification, though.

from arfs.

Hi @ThomasBury, class_weight is needed for multi-class classification, though.

it is, you can use it without any problem for multi-class (Even for binary + imbalance, I'd personally use class_weight, it often leads to better results with a custom weighting vs the class proportional weighting of is_unbalanced)

from arfs.