Data Scientist Interview Prep Guide

A/B-testing questions are one of the most heavily weighted topics in data-science loops, especially for product-DS roles at consumer-tech firms like Meta, Airbnb, and Uber (per DataLemur, founded by ex-Meta/Google data scientist Nick Singh, author of "Ace the Data Science Interview"). Structure it: null/alternative hypotheses; one primary metric (e.g., streaming hours per user per week, a per-user rate that one whale cannot inflate) plus guardrails (churn, content diversity, latency); size from baseline rate, the minimum detectable effect you actually care about, alpha (0.05), and power (0.80) — say those four inputs out loud. Then spend your differentiator time on threats: peeking/early-stopping inflating false positives, multiple-comparison correction, novelty and primacy effects (run at least one full weekly cycle), seasonality, and network/contamination effects from shared accounts. "I’d run it for two weeks at 95% confidence" with no power calc and no threats is the fail.

This is the GenAI question the popular 2026 question banks still skip — and per a 2026 Towards AI survey of GenAI interview questions, "almost every interview I’ve sat in — whether for senior data science, ML engineering, or AI product roles — has shifted toward Generative AI." The verbatim model answer: "LLMs have a knowledge cutoff and can hallucinate on specific facts. RAG grounds generation in retrieved documents, combining the LLM’s language ability with real-time or domain-specific knowledge." Name the four core components from the same source — (1) a document ingestion pipeline with chunking and embedding, (2) a vector store for similarity search, (3) a retriever, and (4) the LLM generator that synthesizes a response from retrieved context. If your loop touches an AI product, prepare this even though the listicles do not list it.

The trade-off articulation here is the whole signal. Per the same 2026 Towards AI source, verbatim: "Use RAG when knowledge needs to be updatable, auditable, or domain-specific without retraining. Fine-tune when: (1) you need consistent output format or style, (2) the task requires skills not in the base model, (3) latency is critical and you can’t afford a retrieval step, or (4) you have 1,000+ high-quality labeled examples." The senior tell is the closing line the source gives: "Often the best architecture combines both: fine-tune for format/style, RAG for factual grounding." Treat it like any DS design question — state the decision criteria, then name the hybrid.

A high-frequency 2026 follow-up once fine-tuning comes up. Verbatim from the Towards AI 2026 set: "Low-Rank Adaptation (LoRA) freezes the pretrained model weights and injects trainable rank-decomposition matrices into each transformer layer. This reduces trainable parameters by ~10,000x compared to full fine-tuning, enabling fine-tuning of 7B+ models on a single GPU." You do not need to derive the math; you need to explain the mechanism (freeze base weights, train small injected matrices) and the practical payoff (drastically fewer trainable parameters, single-GPU feasibility). This is exactly the kind of question that separates a candidate who reads 2022 listicles from one who has worked on a 2026 AI product.

Do not open a query editor first; the interviewer is scoring your decomposition. Step 1: validate the metric itself — is the drop even real, or a logging deploy, a pipeline bug, a dashboard error, or a definition change? Step 2: decompose (session duration = number of actions × time per action) and segment by platform, geography, app version, tenure, and acquisition channel to localize it. Step 3: line it up against deploy logs, feature flags, and a seasonality/marketing calendar. Step 4: form one or two ranked hypotheses and state how you would confirm each. Close with a one-sentence narrative for the PM, not a list of queries. This diagnostic muscle is the mirror image of the A/B-design question — both reward structured skepticism over a memorized recipe.

SQL appears in nearly every data-science loop — do not treat it as the analyst’s problem. Use a LEFT JOIN from January-active users to February-active users and filter WHERE the February side IS NULL, or NOT EXISTS; GROUP BY user and SUM the revenue. Then volunteer the things a data scientist is expected to surface: define "active" precisely (any event? a core action?), handle timezone boundaries on the date filter, and mention partition pruning on the date column and indexing for scale. Saying which approach you chose and why (anti-join vs NOT EXISTS readability/performance) is the judgment signal.

Go past the textbook line. Make it concrete: a deep decision tree memorizing training data (high variance, low bias) versus a linear model that cannot capture a non-linear relationship (high bias, low variance). Then connect to what you actually did — cross-validation for model selection, L1/L2 regularization, and ensembles that cut variance (bagging) or bias (boosting). Close on the production implication: you often accept a slightly higher-bias, lower-variance model because it generalizes and is cheaper to serve and monitor. Interviewers reward the candidate who ties the trade-off to a deployment decision, not just a definition.

This is the ML-depth case that distinguishes a data scientist from an analyst, so show the full arc. Define churn operationally (no trips in 30 days). Identify sources (trip history, earnings trajectory, ratings, support tickets, competitor pricing) and engineer features that encode trend, not just level (trip-frequency slope, earnings trajectory, widening gap between trips). Start interpretable (logistic regression) and justify before reaching for gradient boosting. Evaluate with precision-recall AUC, not ROC AUC, because churn is class-imbalanced — naming that is a senior tell. Then close the loop the interviewer actually cares about: tie each risk tier to a business action (a targeted incentive), and quantify the projected impact in dollars or retained drivers, not in AUC points.

L1 (Lasso) adds an absolute-value penalty and drives coefficients to exactly zero, giving automatic feature selection and a sparse model; L2 (Ridge) adds a squared penalty and shrinks all coefficients proportionally without zeroing them. Choose L1 when you suspect many features are irrelevant and want sparsity for interpretability; choose L2 when you believe most features contribute and you want stable, well-conditioned estimates (especially under multicollinearity). Mention Elastic Net as the blend, and the geometric intuition (diamond vs circular constraint) if the interviewer pushes. The signal is mapping the math choice to a modeling goal, the same way the GenAI round rewards mapping fine-tune-vs-RAG to a product goal.

The clean one-sentence definition: a p-value is the probability of observing an effect at least this large if there were truly no effect. The trap is the common misstatement — it is NOT the probability the null hypothesis is true, and a non-significant result is not proof of no effect. For the misuse half, pull from real experiment practice: peeking and stopping the test the moment it crosses 0.05 (inflating false positives), running many metrics without multiple-comparison correction, or confusing statistical significance with practical (business) significance. Strong statistical fundamentals on p-values, power, and confidence intervals are dealbreakers in the experimentation round at consumer-tech firms.

Use STAR but center how you carried an uncomfortable finding. Give the business context, your methodology, the surprising insight, and how you pressure-tested it (confidence intervals, alternative explanations, a validation check) before you presented. Show you offered an actionable recommendation alongside the bad news rather than just delivering it, and quantify the impact of the decision that followed. The double bind of this role shows up here too: you needed the statistical rigor to trust the result AND the business translation to move a skeptical room — demonstrate both.