Data Scientist Interview Questions & Answers (2026)

Overfitting: model learns training data too well including noise, performs poorly on new data (high variance). Underfitting: model too simple to capture the pattern (high bias). The bias-variance tradeoff: more complexity reduces bias but increases variance. Fix overfitting: regularization (L1/L2), dropout, early stopping, more data, cross-validation.

The CLT states that the distribution of sample means approaches a normal distribution as sample size increases (n ≥ 30), regardless of the population distribution. It justifies using z-tests and t-tests on non-normal data, underpins A/B testing statistics, and means we can make inferences about populations from samples.

(1) Resampling — oversample minority (SMOTE) or undersample majority. (2) Class weights — class_weight='balanced' in sklearn. (3) Threshold tuning — move decision threshold below 0.5 to improve recall. (4) Right metric — use F1, precision-recall AUC, not accuracy. (5) BalancedRandomForest or EasyEnsemble.

Precision = TP/(TP+FP): of predicted positives, how many are correct? Recall = TP/(TP+FN): of actual positives, how many did you catch? There's a tradeoff. F1 = harmonic mean. Choose: maximize recall when false negatives are costly (cancer screening); maximize precision when false positives are costly (spam filtering).

(1) Define hypothesis and primary metric. (2) Power analysis for sample size (80% power, 95% confidence, minimum detectable effect). (3) Random user assignment. (4) Run for a full business cycle. (5) Check for novelty effects and sample ratio mismatch. (6) Two-sided t-test or z-test. (7) Check guardrail metrics. (8) Decision requires statistical AND practical significance.

Random forests build trees independently and average predictions (bagging — reduces variance). Gradient boosting builds trees sequentially, each correcting residual errors of the previous (boosting — reduces bias). XGBoost/LightGBM are optimized implementations. Random forests: faster, less prone to overfitting. Gradient boosting: typically higher accuracy with tuning.

SELECT category, product, revenue FROM (SELECT category, product, SUM(amount) as revenue, RANK() OVER (PARTITION BY category ORDER BY SUM(amount) DESC) as rank FROM orders JOIN products USING(product_id) GROUP BY category, product) ranked WHERE rank <= 3. Key concept: RANK() window function with PARTITION BY.

(1) Collaborative filtering: find users with similar history (user-user or item-item CF), scalable with matrix factorization (ALS, SVD), cold-start problem with new users. (2) Content-based: recommend similar items based on features, no cold-start. (3) Hybrid: combine both. For production: candidate generation (FAISS for ANN) + ranking layer (LightGBM with context features).

Running multiple statistical tests until one reaches p<0.05 by chance. With 20 tests at α=0.05, expect one false positive. Prevention: pre-register hypothesis; Bonferroni correction (divide α by number of tests); use false discovery rate (Benjamini-Hochberg); stop only at pre-calculated sample size; treat initial findings as hypothesis-generating only.

Start with output: 'This model predicts which users will cancel in the next 30 days.' Explain inputs and key drivers using SHAP values: 'Users inactive for 14+ days are 3x more likely to churn.' Avoid jargon. Acknowledge uncertainty: 'It's right about 85% of the time.' Focus on what decisions it enables and the cost of errors.