subtopic

Imbalance strategy preview (train only)

Compare no strategy vs class_weight='balanced' vs SMOTE on the same validation folds; pick ONE. Concepts: class_weight; SMOTE/SMOTENC; resampling only inside training Project A: Your run: class_weight Bad recall 0.72 vs SMOTE 0.63 vs none 0.50 at similar AUC (~0.80) → class_weight alone is the simpler, better choice here. Project B: Not applicable (regression). Equivalent: sample weights to emphasise peak hours, if the cost of peak errors is higher.