Commit Description:

Replaced the data_binary fixture that filtered classes from a multiclass dataset with a new fixture generating a synthetic binary classification dataset using make_classification. This ensures consistent data characteristics, introduces label noise, and better simulates real-world classification challenges.

PR Description:

Summary of Changes:

This PR refactors the data_binary fixture in the test_roc_curve_display.py file. The previous fixture filtered a multiclass dataset (Iris) to create a binary classification task. However, this approach resulted in AUC values consistently reaching 1.0, which does not reflect real-world challenges.

The new fixture utilizes make_classification from sklearn.datasets to generate a synthetic binary classification dataset with the following characteristics:

• 200 samples and 20 features.
• 5 informative features and 2 redundant features.
• 10% label noise (flip_y=0.1) to simulate real-world imperfections in the data.
• Class separation (class_sep=0.8) set to avoid perfect separation.

These changes provide a more complex and representative dataset for testing the roc_curve_display function and other related metrics, thereby improving the robustness of tests.

Reference Issues/PRs:

• Fixes Use more complex data in test_roc_curve_display.py #31243
• See also ENH add from_cv_results in RocCurveDisplay (single RocCurveDisplay) #30399 (comment)

For Reviewers:

• This change ensures that the dataset used for testing is more reflective of real-world data, particularly in classification tasks that may involve noise and less clear separation between classes.