Data Scientist Interview Questions (2026)

Define both error sources, then connect to model complexity: simple models underfit (high bias), complex models overfit (high variance). Tie to concrete levers — regularization, more data, cross-validation — and total error decomposition.

Define it as P(data this extreme | null true). The 'not' is key: it is not the probability the hypothesis is true, and significance ≠ practical importance. Mentioning common misinterpretations is the signal.

Start with the metric (accuracy is misleading; use precision/recall, F1, AUC-PR). Then techniques: resampling (SMOTE, undersampling), class weights, threshold tuning. Stress evaluating on the real distribution.

Contrast bagging (parallel, variance reduction, robust to overfitting) with boosting (sequential, bias reduction, higher accuracy but more tuning-sensitive). Mention training cost and interpretability trade-offs.

Use a window function: DENSE_RANK() OVER (PARTITION BY dept ORDER BY salary DESC), filter rank = 2. Discuss the tie-handling difference between RANK, DENSE_RANK, and ROW_NUMBER.

Distinguish data errors from genuine extremes. Methods: IQR, z-score, domain thresholds, visualization. The judgment call to verbalize — removing real signal can bias the model; investigate before deleting.

Define hypothesis, primary metric, and guardrails. Compute sample size from MDE and power. Address randomization unit, novelty effects, and when to stop. Pre-register the success criterion.

Investigate distribution shift, train/serve skew, leakage in offline data, and feedback loops. Propose monitoring and a shadow deployment. The systems thinking is what's being tested.

L1 (Lasso) induces sparsity and does feature selection; L2 (Ridge) shrinks coefficients smoothly. Explain geometrically why L1's corners zero out weights, and when each is preferable.

STAR. Show the question, your method, and crucially the communication that drove action. Quantify the outcome. Data scientists who can't influence decisions don't get hired.