Sign in Sign up

Interview questions · Data Scientist

Data Scientist interview questions

Recruiters hiring data scientists want candidates who can translate ambiguous business problems into rigorous analytical frameworks, build and validate predictive models, and communicate findings clearly to non-technical stakeholders. They assess statistical depth, coding proficiency (Python or R), and the ability to deliver actionable insights under real-world constraints. Commercial awareness and the ability to work with messy, incomplete data are equally valued alongside academic rigour.

Walk me through a model you built end-to-end, from problem definition to deployment.

Recruiters want to see the full data science lifecycle — not just modelling — including how you scoped the problem and handled productionisation.

Model answerI built a customer churn model for a subscription business with 200k accounts. I defined churn as no activity within 30 days, engineered 40 features from clickstream and billing data, and compared logistic regression, gradient boosting, and a random forest before selecting XGBoost (AUC 0.87). I packaged the model as a FastAPI microservice, deployed it on AWS Lambda, and wired it to the CRM so the retention team received daily ranked call lists. Churn in the targeted segment dropped by 11% over the following quarter.

How do you handle class imbalance in a binary classification problem?

Imbalanced datasets are common in real-world data science; this tests practical statistical knowledge beyond textbook scenarios.

Model answerI first check whether the imbalance is informative — sometimes the minority class is genuinely rare and should stay rare in the model. Then I choose between oversampling the minority class with SMOTE, undersampling the majority, or adjusting class weights in the loss function. I always evaluate with PR-AUC or F1 rather than accuracy, since a naive classifier that predicts the majority class can score 97% accuracy while being useless. On a recent fraud detection task SMOTE combined with class weighting lifted recall on the positive class from 51% to 78% with only a 4-point precision drop.

Describe a situation where your analysis contradicted a senior stakeholder's intuition. How did you handle it?

Data scientists must influence decisions without authority; this tests communication and persuasion under pressure.

Model answerOur CMO was convinced that a particular email campaign was our top revenue driver. My attribution analysis using a Shapley-value model showed it was actually cannibalising organic conversions. I presented the finding in a one-pager with clear confidence intervals and a proposed A/B test to settle the question. The CMO agreed to run the test; the holdout group showed no revenue difference, and we reallocated 80k of campaign budget to a higher-ROI channel.

How do you validate that a model is not leaking future information (data leakage)?

Data leakage produces over-optimistic results in development and fails silently in production — a costly mistake.

Model answerI enforce strict temporal splits for any time-series or event-based problem: training data ends before the earliest timestamp in the validation set, with no look-ahead features. I also audit every feature's derivation to check whether it could encode the target label — for example, a 'days since last complaint' feature computed using the complaint resolution date would leak. I run a simple sanity check: if a baseline model with only time-invariant features dramatically underperforms the full model, that gap warrants investigation.

Tell me about a time you had to work with severely incomplete or low-quality data.

Real-world data is rarely clean; recruiters want to see pragmatic problem-solving and transparency about uncertainty.

Model answerI was asked to forecast weekly demand for 3,000 SKUs, but 40% of the historical records had missing sales figures due to a legacy ERP migration. I identified that the missingness was not random — it correlated with a warehouse system outage — so I treated those periods as censored rather than zero-sales, imputed using a seasonal median, and flagged those SKUs with a confidence band in the output. Forecast accuracy (MAPE) was 14%, versus a naive baseline of 23%, and the supply chain team could visually identify lower-confidence SKUs.

Explain a time you communicated a complex statistical concept to a non-technical audience.

Data scientists who cannot translate findings into business language have limited organisational impact.

Model answerI needed to explain confidence intervals to a board who kept asking for a single-point revenue forecast. I used the analogy of a weather forecast — we do not say it will rain 17.4mm; we say expect 10-25mm. I built a simple visual showing three scenarios (pessimistic, central, optimistic) rather than probability distributions. The board adopted scenario planning, which led to a more resilient inventory strategy that avoided a 200k stockout the following quarter.

How do you monitor a deployed model for performance degradation over time?

Models decay as data distributions shift; operational monitoring is a key part of production data science.

Model answerI instrument every deployed model with two layers: prediction monitoring (score distribution, prediction drift via PSI) and outcome monitoring (comparing predicted vs actual labels on a rolling 30-day window). I set alerting thresholds — if PSI exceeds 0.2 or AUC drops more than 3 points below baseline, a Slack alert fires and triggers a manual review. On a pricing model I managed, this caught a feature drift caused by a supplier data feed change within 48 hours, preventing an estimated 40k in mispriced orders.

Describe your approach to feature engineering for a dataset with high-cardinality categorical variables.

High-cardinality categoricals (e.g., product IDs, user IDs) break naive one-hot encoding and test advanced preprocessing knowledge.

Model answerFor a product recommendation model with 50,000 SKU codes, I used target encoding with leave-one-out smoothing to avoid overfitting on rare products, cross-validated the encoding inside each fold to prevent leakage. For very rare categories (fewer than 10 occurrences) I grouped them into an other bucket. I also tested embedding layers in a neural network approach, which performed comparably but at higher inference cost — so we shipped the target-encoded gradient booster for latency reasons.

Tips for this role

Practise for THIS role

Paste your job offer, pick a recruiter, and run a mock interview with a detailed coaching report.

🎤 Start an interview

Everything you can do

Simulated interview

Practice out loud with an AI recruiter, in real conditions.

7 AI recruiters

From the friendly HR to the tough manager: pick your challenge.

ATS-ready CV

Generate a clean PDF CV that gets past automated filters.

Video interview

Facing a realistic video avatar: also work on your posture and presence.

Coaching report

A score out of 100, your strengths, areas to improve and quotes from your answers.

Personality analysis

Assess your Big Five traits with the IPIP-NEO test and get a summary to share with HR.

CV / job match

Measure how well your CV matches a job before applying.

Cover letter

A tailored cover letter, matched to the job and your CV, ready to send as PDF.

Prep brief

Likely questions, angles to prepare and recent info about the company.

Tailored CV for the offer

Your CV reworked for each offer: skills and experience prioritized to match what the role asks for.

Model answers (STAR)

Your interview questions rewritten as exemplary STAR answers.

Thank-you email

The perfect follow-up message, based on your interview strengths.

Retake & improve

Replay the interview as much as you like: each report tracks your progress.