Data Scientist & Machine Learning Interview Questions

When dealing with extreme class imbalance like credit card fraud, first eliminate raw accuracy as an evaluation metric and switch to Precision-Recall AUC (PR-AUC) or F-beta score depending on the business cost of false negatives. At the algorithm level, I implement cost-sensitive learning by adjusting class weights in the loss function (or using Focal Loss). At the data level, rather than naive random oversampling, I use SMOTE with Tomek links or treat the problem as an unsupervised anomaly detection task using Isolation Forests or autoencoders.

The bias-variance tradeoff reflects the tension between model simplicity (bias) and flexibility (variance). High bias causes underfitting, while high variance causes the model to memorize training noise, leading to poor generalization. Regularization constrains model complexity: L2 regularization (Ridge) adds the squared sum of coefficients to the loss function, shrinking weights toward zero and distributing influence smoothly. L1 regularization (Lasso) adds the absolute sum, driving non-essential coefficients completely to zero, functioning as automated feature selection.