Repeated measurements from each patient are randomly split into folds. Why may the score exaggerate performance on new…
Classic MLLevel: IntermediateGroup-disjoint evaluation
Repeated measurements from each patient are randomly split into folds. Why may the score exaggerate performance on new patients?
Hints
- Identify the independent sampling unit.
- The deployment target is a new person.
Answer
Patient-specific signals can appear on both sides. GroupKFold keeps each patient entirely on one side, evaluating transfer to unseen patients rather than recognizing known individuals.
Source and licence: scikit-learn 1.7.2 — cross validation · BSD-3-Clause · answer: Quantoros editorial