Data Scientist Interview Prep Guide

Structure your answer: state null and alternative hypotheses clearly, identify the north star metric (e.g., streaming hours per user per week) and guardrail metrics (churn rate, content diversity). Calculate sample size using baseline conversion, MDE of 2%, and 80% power. Discuss randomization unit (user-level), test duration (at least 2 weeks to capture weekly patterns), and how you handle network effects. Mention novelty effect and long-term holdout groups.

Do not jump to solutions. First clarify: is this across all users or specific segments? Decompose the metric: session duration = number of actions x time per action. Segment by platform, geography, user tenure, and feature. Check for data pipeline issues, recent releases, app crashes, and seasonal effects. Build a hypothesis tree, prioritize by data availability, and propose 3 next steps with expected findings.

Use a LEFT JOIN from January active users to February active users and filter WHERE February is NULL, or use NOT EXISTS. Group by user and SUM revenue. Mention potential optimizations: partition pruning on date columns, appropriate indexing. Discuss how you would handle timezone issues and define "active" precisely.

Go beyond the textbook definition. Use a concrete example: a decision tree with max depth overfitting training data (high variance, low bias) vs a linear model underfitting non-linear relationships (high bias, low variance). Connect to practical solutions: cross-validation for model selection, regularization (L1/L2), ensemble methods that reduce variance (bagging) or bias (boosting). Show you understand the implications for model deployment.

Structure: define churn (no trips in 30 days), identify data sources (trip history, earnings, ratings, support tickets, competitor pricing). Feature engineering: trip frequency trend, earning trajectory, peak-hour participation, gap between trips. Model choice: start with logistic regression for interpretability, then gradient boosting for performance. Evaluation: use precision-recall AUC (not ROC AUC due to class imbalance). Discuss deployment: real-time scoring for intervention triggers, business action tied to each risk tier.

L1 (Lasso) adds absolute value penalty, drives coefficients to exactly zero, performing automatic feature selection. L2 (Ridge) adds squared penalty, shrinks all coefficients proportionally. Use L1 when you suspect many features are irrelevant and want a sparse model. Use L2 when all features likely contribute and you want stable coefficient estimates. Discuss Elastic Net as a combination, and mention the geometric interpretation (diamond vs circle constraint regions).

Feature engineering is key: transaction velocity, geographic anomalies, device fingerprinting, time-of-day patterns, amount deviation from user baseline. Model pipeline: rule-based system for obvious fraud, gradient boosting for scoring, anomaly detection (Isolation Forest) for novel patterns. Address class imbalance: SMOTE, cost-sensitive learning, or threshold optimization. Discuss the precision-recall tradeoff: blocking legitimate transactions costs revenue, missing fraud costs trust. Mention real-time vs batch scoring needs.

Use STAR but emphasize how you communicated uncomfortable findings. Describe the business context, your rigorous methodology, the surprising insight, and how you presented it with confidence intervals and alternative explanations. Show you validated the finding before presenting, offered actionable recommendations alongside the bad news, and ultimately influenced the decision. Quantify the business impact of acting on your analysis.