Class Imbalance

Class imbalance means highly skewed class ratios (e.g. 1:1000). Problem and impact: the model favors the majority class and the decision boundary is dragged toward the minority; accuracy becomes meaningless — predicting the majority class always yields 99.9% accuracy while minority recall collapses toward zero, yet the business value lives in the minority (fraud, disease). Metric choice: when the minority class matters, PR curves / PR-AUC beat ROC — ROC's FPR denominator contains a huge number of true negatives that dilute the curve into over-optimism, whereas PR directly captures the precision-recall (minority) tradeoff; pair with F1, F-beta and PR-AUC. The solution map has four layers: data (SMOTE oversampling, undersampling, synthetic data), algorithm (class weights / cost-sensitive loss, Focal Loss), metrics (PR, F1, PR-AUC), and decision (threshold moving).

fraud detection, anomaly detection, rare-disease diagnosis, low-conversion ads — tasks at 1:100 or even 1:10000 ratios; interviews always ask “how to handle a 1:100 ratio” and “why not look at accuracy”.

naive training makes the model lazily predict the majority class, yielding high accuracy but zero utility; the correct order is to fix the metric first (PR over ROC), then intervene at the data, algorithm and decision layers (resampling + weighted loss + threshold moving) to pull minority recall back to a business-usable level.

Why does PR beat ROC under class imbalance? How does the huge true-negative denominator dilute FPR and thus ROC?

Positive:negative = 1:1000, 99.9% accuracy, but business needs high recall: what are your steps (metric → method → validation)?

How does class weight (cost-sensitive learning) work? Is it mathematically equivalent to oversampling?

How do you move the decision threshold? How does it relate to resampling and weighted loss?

Under imbalance, which reflects business better, F1 or AUC? When is AUC still fine?