Class Imbalance
When fraud is one case in a thousand, a model that never flags fraud is 99.9% accurate. Imbalance is a fact to evaluate around, not a defect to remove.
-
-
Stratified Group K-Fold
A chance-level problem scored 0.9198 when groups were split across folds and 0.5323 when they were kept together. The truth was 0.5000.
-
Imbalanced Data
Rare outcomes raise distinct questions about fitting, decision thresholds, and evaluation; resampling is one possible training intervention.