Data Science Interview Help — AI for ML, Stats, SQL & Case Rounds

1. Coding & SQL screen

Python data manipulation (pandas, NumPy), a LeetCode-style algorithm, and almost always SQL: joins, window functions, aggregations, and cohort/retention queries. The AI returns a working query or function with complexity notes — you still explain the approach aloud.

2. Statistics & probability

Hypothesis testing, p-values, confidence intervals, Bayes, distributions, and the classic experiment-design questions. Expect "explain a p-value to a PM" and "how would you detect if this metric change is real?" The AI scaffolds a crisp, correct definition you can deliver without rambling.

3. ML theory & modeling

Bias/variance, regularization (L1 vs L2), tree ensembles vs linear models, evaluation metrics (precision/recall, AUC, log-loss), handling class imbalance, and overfitting. The AI maps the question to the trade-off the interviewer is probing.

4. ML system design

"Design a recommendation system / fraud detector / feed ranker." Graded on data, features, model choice, training/serving split, evaluation, and online metrics. The AI lays out the standard skeleton (problem framing → data → features → model → offline/online eval → deployment & monitoring) so you don't miss a stage.

5. Product / metrics case

"Pick a metric for X." "DAU dropped 5% — investigate." A/B test design, guardrail metrics, and trade-off reasoning. The AI surfaces a structured root-cause tree and reminds you about novelty effects, seasonality, and sample size.

• "Explain the bias–variance tradeoff." — High bias underfits (model too simple, misses signal); high variance overfits (model too flexible, fits noise). Total error decomposes into bias squared plus variance plus irreducible error, and you tune model complexity to minimize their sum on held-out data.
• "How do you handle overfitting?" — Regularization, more data, and simpler models. Be specific: L1 (Lasso) adds the absolute-value penalty and drives some coefficients to exactly zero, doubling as feature selection; L2 (Ridge) adds the squared penalty and shrinks coefficients smoothly without zeroing them. Mention early stopping, dropout, and cross-validation too.
• "When does accuracy mislead, and what do you use instead?" — On imbalanced classes (e.g. 1% fraud), a model predicting "never fraud" scores 99% accuracy yet is useless. Use precision, recall, and their harmonic mean F1, plus ROC-AUC — and prefer the precision–recall curve when positives are rare.
• "Walk me through train/validation/test splits and k-fold cross-validation." — Train fits parameters, validation tunes hyperparameters, test gives an unbiased final estimate touched only once. k-fold rotates the validation slice across k partitions and averages, which uses data efficiently and reduces variance in the estimate; split before any fitting to avoid leakage.
• "How does an A/B test decide significance?" — State a null and alternative hypothesis, pick α (commonly 0.05) and power (commonly 0.80), and compute the required sample size up front. The p-value is the probability of a result this extreme if the null were true — it is not the probability the null is true. Watch for peeking, multiple testing, and novelty effects.
• "Supervised vs unsupervised — and how do you treat missing data and class imbalance?" — Supervised learns from labeled targets (regression, classification); unsupervised finds structure without labels (clustering, dimensionality reduction). For missing data, distinguish the mechanism (MCAR/MAR/MNAR) and choose deletion vs imputation accordingly; for imbalance, use resampling (SMOTE), class weights, and threshold tuning rather than raw accuracy.
• A SQL or probability question — e.g. "find the second-highest salary per department" (a window function like DENSE_RANK() OVER (PARTITION BY dept ORDER BY salary DESC)), or "what's the probability of at least one shared birthday in a room of 23?" Talk through the approach aloud; the interviewer grades reasoning, not just the final value.