Top Data Scientist Interview Questions & Answers

The two sources of error: • Bias — Error from oversimplified assumptions (underfitting). A linear model fitting nonlinear data has high bias. • Variance — Error from sensitivity to training data (overfitting). A deep decision tree that memorizes noise has high variance. The tradeoff: Reducing bias (adding complexity) increases variance and vice versa. The sweet spot minimizes total error = bias² + variance + irreducible noise. Practical guidance: 1. Start simple (high bias, low variance) 2. Increase complexity until validation error stops improving 3. Use regularization (L1/L2) to manage the tradeoff directly 4. Ensemble methods like Random Forests reduce variance without increasing bias by averaging many high-variance models

Random Forest (bagging): • Trains trees independently, averages predictions • Robust, parallelizable, hard to overfit • Use when: fast baseline needed, noisy data, interpretability via feature importance Gradient Boosted Trees (boosting): • Trains sequentially — each tree corrects previous errors • Typically higher accuracy but more overfitting risk • Use when: maximizing predictive performance, clean data, time to tune hyperparameters Quick decision guide: • Need a quick, stable model? → Random Forest • Optimizing for competition/production accuracy? → XGBoost/LightGBM • Dataset is small or noisy? → Random Forest • Dataset is large and clean? → Gradient Boosting

Four parameters determine sample size: 1. Baseline conversion rate — your current metric 2. Minimum detectable effect (MDE) — smallest improvement worth detecting 3. Significance level (α) — usually 0.05 4. Statistical power (1-β) — usually 0.80 Formula: n = (Z_α/2 + Z_β)² × (p1(1-p1) + p2(1-p2)) / (p1-p2)² Key insight — smaller effects need exponentially more data: • Baseline 10%, detect 1% absolute lift → ~14,000 users/group • Baseline 10%, detect 0.5% absolute lift → ~57,000 users/group Practical tip: Choose a realistic MDE before the test starts — it directly determines how long you'll run it. Use a power analysis calculator rather than computing by hand.