The 'group' parameter cannot be applied in the roles of the fit_predict function in TabularAutoML.

[yuuniee]

🐛 Bug

• The 'group' parameter cannot be applied in the roles of the fit_predict function in TabularAutoML.
• As a result of backtracking the code, I found that the 'group' item was not included in the CV split process.

To Reproduce

Steps to reproduce the behavior:

1. Run -> TabularAutoML().fit_predict(..., roles={'target': TARGET_NAME, 'group': GROUP_NAME})
2. "lightautoml > reader > utils.py > set_sklearn_folds()" <- Here, check if the 'group: Optional[np.ndarray]' variable has a value.

Expected behavior

• The CV split method changes depending on the 'group' designation above.

Additional context

• This was resolved by adding the following code to line 322 of "lightautoml > reader > base.py".
• "kwargs["group"] = train_data.loc[:, roles["group"]]"

The issue was found in version "0.3.8.1" and is unknown for other versions.

I am always grateful for your open source contributions.
Thank you.

Checklist

• bug description
• steps to reproduce
• expected behavior
• code sample / screenshots

[alexmryzhkov]

Hi @yuuniee,

Thanks for the detailed description of the problem with the models validation. We will for sure take a look into it.

Alex

[screengreen]

Hi @yuuniee,

We checked correctness of CV split with the 'group' parameter. We tried several most popular task for it ['binary', 'multiclass', 'reg', 'multilevel'] and didn't find any issues.

what you tried to add "kwargs["group"] = train_data.loc[:, roles["group"]]" is implemented here:

LightAutoML/lightautoml/reader/base.py

Line 311 in b4ace0c

kwargs[attrs_dict[r.name]] = train_data.loc[:, feat]

if you want to examine it yourself, you can check the attached .ipynb file, that shows the 'group' parameter effect.
Tutorial_1_basics_check_group_parameter.ipynb.zip

if it doesn't solve your problem, please provide additional information

Andrew