Handle imbalance, sampling, and overfitting

## Practical ML questions (classification and generalization) Answer the following ML engineering/data science questions. ### A) Class imbalance You’re training a classifier where the positive class is rare. - How do you handle **class imbalance** (data-level and algorithm-level approaches)? - Which **evaluation metrics** are appropriate and why (e.g., accuracy vs precision/recall/F1/ROC-AUC/PR-AUC)? - What pitfalls should you watch for (e.g., calibration, thresholding, leakage)? ### B) Training on a sample from a very large dataset You train a model on a sample drawn from a massive dataset. - How do you verify the **sample is representative**? - How do you validate that a model trained on the sample will **generalize** to the full population? ### C) Preventing overfitting in tree-based models For decision trees / random forests / gradient-boosted trees: - What knobs and practices help prevent **overfitting**? ### D) Why L1/L2 regularization is biased Explain why L1 (Lasso) and L2 (Ridge) regularization typically produce **biased** coefficient estimates, and why we still use them.