[python-package] adding max_category_values parameter to create_tree_digraph method (fixes #5687)

[moziada]

fixes #5687

[jmoralez]

Thanks for your contribution! I've left some minor comments

[moziada]

I can's see why R-package checks are failing

[jameslamb]

please merge laster master into this branch to get the changes from #5821

[jmoralez]

LGTM. Thanks a lot for your contribution!

[jmoralez]

@jameslamb do you want to review this as well?

[jameslamb]

Thanks very much for this!

Can you please just add unit tests covering this functionality? That would give us more confidence that this is working and prevent it from being broken in the future.

Please add two tests, both with datasets that have at least one categorical feature that is informative and used in splits.

• one where the len(category_values) > max_category_values condition you've added is True
• one where that condition is false

You could, for example, copy this test:

LightGBM/tests/python_package_test/test_plotting.py

Line 160 in ffb2986

def test_create_tree_digraph(breast_cancer_split):

[moziada]

After looking at sklearn datasets I did not find any classification dataset that contains categorical features, I am thinking of creating multiple base datasets with sklearn.datasets.make_classification each one has different distribution then merging them into one dataset and use the label of each base dataset as a categorical feature. what do you think?

[jmoralez]

I think that could work, you just need to make sure that the target depends on those features somehow (so that they're chosen for some splits), maybe something like what we do here

LightGBM/tests/python_package_test/test_dask.py

Lines 185 to 205 in 5989405

	if output == 'dataframe-with-categorical':
	num_cat_cols = 2
	for i in range(num_cat_cols):
	col_name = f"cat_col{i}"
	cat_values = rnd.choice(['a', 'b'], X.shape[0])
	cat_series = pd.Series(
	cat_values,
	dtype='category'
	)
	X_df[col_name] = cat_series
	X = np.hstack((X, cat_series.cat.codes.values.reshape(-1, 1)))
	# make one categorical feature relevant to the target
	cat_col_is_a = X_df['cat_col0'] == 'a'
	if objective == 'regression':
	y = np.where(cat_col_is_a, y, 2 * y)
	elif objective == 'binary-classification':
	y = np.where(cat_col_is_a, y, 1 - y)
	elif objective == 'multiclass-classification':
	n_classes = 3
	y = np.where(cat_col_is_a, y, (1 + y) % n_classes)

[moziada]

I have created a quantized version of breast cancer dataset and used the features as a categorical features. I made only one test case where the condition is false (categorical values should not be compressed), but the problem is the way that lightgbm splits on categorical features is not deterministic, for example a feature could have 30 different categorical values but a random indexed tree may be only splits on one categorical value. what do you suggest and I need your review

[jameslamb]

Nice idea binning the breast_cancer dataset to create an informative dataset of categorical features! I think that's a great approach for this PR's tests.

Please see my suggestions on how to proceed, and please add at least one test where the number of categories in a feature is greater than the value of max_category_values as well.

[jameslamb]

@moziada sorry for the failing R CI jobs...those are not related to your PR. We've recently fixed that issue in #5859 .

Can you please merge latest master into this branch? Once you do that, I'll review the testing changes you've pushed.

[moziada]

Are there any updates? @jameslamb

[jameslamb]

Looks good to me, thanks for the contribution!

[github-actions]

This pull request has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.