The starting point is scikit-learn's Wisconsin Diagnostic Breast Cancer dataset.
Source documentation: https://scikit-learn.org/stable/modules/generated/sklearn.datasets.load_breast_cancer.html
Target encoding in scikit-learn is 0 = malignant and 1 = benign. For this experiment I create a controlled imbalance from the original observations:
- all 212 class-0 (malignant) rows are kept;
- 35 class-1 (benign) rows are selected with a seeded random sample;
- no synthetic rows are generated.
The resulting dataset is intentionally artificial in its class balance. It is useful for studying class weighting, but it should not be interpreted as a realistic estimate of medical prevalence.