Top Data Scientist Interview Questions & Answers

Bias is error from incorrect model assumptions — a high-bias model underfits and performs poorly on training and test data. Variance is sensitivity to noise in the training data — a high-variance model overfits and performs well on training data but poorly on test data. The tradeoff: increasing model complexity reduces bias but increases variance. The goal is the minimum total error, achieved by selecting the right model complexity and regularization.

Options: (1) Resampling — oversample the minority class (SMOTE) or undersample the majority class. (2) Class weights — penalize misclassification of the minority class more heavily in the loss function. (3) Threshold tuning — adjust the decision threshold based on the precision-recall tradeoff for your business objective. (4) Ensemble methods — boosting algorithms (XGBoost, LightGBM) handle imbalance well. Evaluate with AUC-ROC or precision-recall, not accuracy.

Regularization adds a penalty to the loss function to reduce model complexity and prevent overfitting. L1 (Lasso) adds the absolute value of coefficients — it drives some weights to exactly zero, performing feature selection. L2 (Ridge) adds the squared value — it shrinks weights toward zero but rarely to exactly zero. Use L1 when you want feature selection; L2 when all features may contribute but need to be constrained.

Accuracy is misleading for imbalanced classes. Use precision and recall for the minority class, F1-score as the harmonic mean, AUC-ROC for ranking performance across thresholds, AUC-PR for imbalanced classification. For regression: RMSE (penalizes large errors), MAE (robust to outliers), MAPE (interpretable as a percentage). Always evaluate on a held-out test set, not the training or validation set.

Define the hypothesis and the primary metric you are testing. Determine the minimum detectable effect and calculate the required sample size (using a power calculation). Randomly assign users to control and treatment groups, ensuring they are representative. Run the experiment until you reach the required sample size — do not stop early based on interim results (p-hacking). Analyze with the pre-specified metric, check for novelty effects and interaction effects, and make a decision with statistical confidence.

Use the STAR format. Be specific about the business question, the data you used, the analysis you ran, the insight you surfaced, and the decision that changed as a result. Quantify the impact if possible. The strongest answers show you communicated the insight clearly to a non-technical stakeholder and drove action.