How would you explain the difference between supervised and unsupervised learning, and when you would use each?

Start by defining both in simple terms, supervised learning predicts a labeled outcome, and unsupervised learning finds structure in unlabeled data. Then explain how you choose between them based on the problem, the data you have, and what decision the business needs to make. For example, if you have historical outcomes like “customer churned yes or no,” you can train a supervised classifier to predict churn risk and drive retention outreach. If you do not have labels and you want to understand customer types, you can use clustering to segment users based on behavior and then test different messaging by segment. Avoid treating unsupervised outputs as truth without validation, clusters are hypotheses that need human interpretation and follow-up metrics. In supervised settings, call out how you handle imbalance, data leakage, and how you translate model scores into an action threshold.

Walk me through how you would evaluate a classification model.

Explain that you start with the decision the model supports, then pick metrics that match the cost of errors. You also describe your evaluation setup, train and validation split, cross-validation when data is limited, and a final holdout test if you can afford it. As a concrete example, for fraud detection you often care more about recall at a fixed false positive rate than raw accuracy. You might report precision, recall, F1, ROC-AUC, and also a precision-recall curve because positives are rare, then pick a threshold that meets investigation capacity. Mention common traps, accuracy is misleading on imbalanced data, and tuning on the test set inflates results. Add how you check calibration if probabilities are used for ranking or expected value, and how you monitor drift after launch so performance does not silently degrade.

What is data leakage, and how do you prevent it?

Define data leakage as using information during training that would not be available when the model is making predictions in production. Then give a clear prevention strategy, design features from the perspective of prediction time, and keep your preprocessing inside the training pipeline. For example, if you predict loan default, a feature like “number of missed payments in the last 30 days” might leak if that window includes days after the loan decision date. Another common leak is fitting a scaler or target encoder on the full dataset before splitting, which lets test information influence training. Prevention comes down to discipline and checks, time-based splits for time series, training-only fit for transforms, and careful joins with point-in-time logic. You can also run sanity tests like training with shuffled labels, if performance stays high, something is probably leaking.

How would you design an A/B test, and what pitfalls would you watch for?

Start with the goal and the primary metric, then define the hypothesis, unit of randomization, and the minimum detectable effect you care about. You also describe how you set guardrail metrics, run a power analysis to pick sample size, and decide in advance when you will stop the test. For example, for an onboarding change you might randomize at the user level, track activation rate as the primary metric, and use retention and support tickets as guardrails. If the change can spill over between users, like a marketplace feature, you might need cluster randomization or to choose a different unit. Call out pitfalls that often break experiments, peeking early, running too many metrics without a plan, and ignoring novelty effects. Also mention sample ratio mismatch, bots, and multiple exposures, plus how you handle users who switch devices or appear in both variants.

Explain bias-variance tradeoff in practical terms.

Frame it as two ways a model can fail, high bias means the model is too simple and underfits, high variance means it is too sensitive to noise and overfits. Then connect it to what you would actually do, change model complexity, add regularization, get more data, or improve features. A practical example is predicting house prices, a simple linear model might miss non-linear patterns like neighborhood effects, which is bias. A deep tree might fit every odd pricing outlier in the training set, which is variance, and you would see train performance far better than validation. Tie it to diagnostics, learning curves, train versus validation gap, and stability across folds. Also mention that feature quality often beats model complexity, and you should confirm gains with cross-validation rather than trusting a single split.

How do you handle missing data, and how do you choose an imputation strategy?

Start by saying you first diagnose why data is missing, missing completely at random, missing at random, or missing not at random, because the cause affects bias. Then explain you choose an approach based on the model, the feature meaning, and how missingness relates to the target. For example, for numeric features you might try median imputation plus a missing indicator, especially for tree-based models that can use the indicator to learn patterns. For time series, forward fill might make sense for sensors, but it could be wrong for transactional data where “no event” is meaningful. Mention that you should not impute using information from the future, and you should fit imputers on training data only. Also call out that sometimes the right answer is to redesign data collection, or to treat missingness as a signal rather than “fixing” it away.

Describe how you would approach a SQL question where you need the top N per group.

Explain that you typically use a window function like ROW_NUMBER or RANK partitioned by the group and ordered by the value you care about, then filter to N. If the database does not support window functions, you can fall back to correlated subqueries, but window functions are usually clearer and faster. As an example, to find the top 3 products by revenue per category, you would compute revenue per product, then apply ROW_NUMBER() OVER (PARTITION BY category ORDER BY revenue DESC) as rn, and select rows where rn <= 3. You should also decide how to handle ties, RANK keeps ties and can return more than N. Common mistakes are filtering before computing the window, or forgetting that joins can duplicate rows and inflate revenue. A good habit is to validate counts at each step, and to run a small sample query to confirm the ranking logic matches the business definition.

How would you choose between logistic regression, random forest, and gradient boosting for a classification problem?

Start by mapping model choice to constraints, interpretability needs, data size, feature types, and how much time you have for iteration. Then explain that you would begin with a simple baseline and only move to more complex models when you can show real improvement on a proper validation setup. For example, logistic regression is a strong baseline, fast, and easier to explain, especially with well-engineered features and regularization. Random forests handle non-linearities with less tuning, while gradient boosting often wins on tabular data but needs careful tuning, monotonic constraints if required, and more attention to leakage and drift. Mention that you compare not just AUC, but also calibration, stability across time, and operational cost like latency and retraining. Also call out that the “right” model depends on how it will be used, a slightly worse model that is explainable and stable can be the better business choice.

How do you detect and handle outliers?

Start by clarifying what “outlier” means in context, a true error, a rare but real case, or a different population. Then explain you would use both visual checks and statistical rules, and you would decide treatment based on downstream impact. For example, in payment data, a negative amount might be a data error that should be fixed or removed, while a very large amount could be a real enterprise customer. You might use IQR rules, z-scores on log-transformed values, or model-based methods, then compare distributions by segment to see if the “outliers” are actually a subgroup. Avoid blindly deleting points, that can hide important behavior and bias the model. Prefer approaches like winsorizing, using more stable loss functions, adding caps based on domain limits, or modeling segments separately when the business logic supports it.

Tell me about a data science project you are proud of, and how you measured success.

Give a tight story, problem, data, approach, and outcome, then connect it to the role you are interviewing for. You should highlight both your technical choices and your judgment, what you tried, what you rejected, and how you knew it worked. For example, you might describe building a churn model, starting with a baseline logistic regression, then adding time-based features and a gradient boosting model. You measured success with lift in the top risk decile, calibrated probabilities for prioritization, and an offline to online check where the retention team tracked saved accounts. Include what you learned and what you would improve, like better point-in-time features, clearer monitoring, or a simpler model for easier adoption. Interviewers listen for ownership and clarity, so avoid jargon and be explicit about how your work changed a decision or process.

Tell me about a time you disagreed with a stakeholder about a model or metric.

Situation: A product manager wanted to ship a recommendation model based on higher click-through rate, but the change increased low-quality clicks. Task: You needed to align on a metric that matched user value and still let the team move quickly. Action: You pulled a sample of sessions and showed that clicks rose mainly from sensational items, while downstream saves and dwell time fell. You proposed a primary metric that combined saves per session with a guardrail on negative feedback, and you offered a quick A/B test plan with a clear stop rule. Result: The team ran the experiment and found CTR up but saves down, so you did not ship that version. You shipped an adjusted model two weeks later that improved saves while keeping CTR flat, and the PM adopted the new metric for future iterations.

Describe a time your analysis was wrong or incomplete. What did you do next?

Situation: You reported that a new onboarding flow improved activation, but a week later support tickets spiked and retention dropped. Task: You had to find the error, correct the story, and rebuild trust with the team. Action: You rechecked the experiment setup and found that your analysis window excluded users who signed up on weekends, which skewed the sample. You reran the analysis with the correct cohort definition, added sensitivity checks, and walked the team through what changed and why. Result: The team rolled back the change and planned a revised test with a longer window and stronger guardrails. Your follow-up checklist became part of the team process, and you started doing peer review on experiment analyses to reduce repeat mistakes.

Tell me about a time you had to explain a complex result to a non-technical audience.

Situation: You built a forecasting model for inventory planning, and the operations team was skeptical because the model output looked different from their spreadsheet. Task: You needed to explain the model in a way that helped them trust and use it. Action: You translated the model into three simple drivers they already cared about, seasonality, promotions, and stockouts, and showed backtests on past months they remembered well. You used a few real product examples to compare forecasts, and you explained uncertainty with prediction intervals and what actions to take when the interval was wide. Result: The team started using the forecast for weekly planning, and they created a feedback channel to flag items where the model missed. Over the next quarter, forecast errors dropped as you added their domain inputs, and adoption improved because they could connect model behavior to real-world events.

data science Interview Questions: Complete Guide

Q: How would you choose between logistic regression, random forest, and gradient boosting for a classification problem?

Start by mapping model choice to constraints, interpretability needs, data size, feature types, and how much time you have for iteration. Then explain that you would begin with a simple baseline and only move to more complex models when you can show real improvement on a proper validation setup. For example, logistic regression is a strong baseline, fast, and easier to explain, especially with well-engineered features and regularization. Random forests handle non-linearities with less tuning, while gradient boosting often wins on tabular data but needs careful tuning, monotonic constraints if required, and more attention to leakage and drift. Mention that you compare not just AUC, but also calibration, stability across time, and operational cost like latency and retraining. Also call out that the “right” model depends on how it will be used, a slightly worse model that is explainable and stable can be the better business choice.

Q: How do you detect and handle outliers?

Start by clarifying what “outlier” means in context, a true error, a rare but real case, or a different population. Then explain you would use both visual checks and statistical rules, and you would decide treatment based on downstream impact. For example, in payment data, a negative amount might be a data error that should be fixed or removed, while a very large amount could be a real enterprise customer. You might use IQR rules, z-scores on log-transformed values, or model-based methods, then compare distributions by segment to see if the “outliers” are actually a subgroup. Avoid blindly deleting points, that can hide important behavior and bias the model. Prefer approaches like winsorizing, using more stable loss functions, adding caps based on domain limits, or modeling segments separately when the business logic supports it.

Q: Tell me about a data science project you are proud of, and how you measured success.

Give a tight story, problem, data, approach, and outcome, then connect it to the role you are interviewing for. You should highlight both your technical choices and your judgment, what you tried, what you rejected, and how you knew it worked. For example, you might describe building a churn model, starting with a baseline logistic regression, then adding time-based features and a gradient boosting model. You measured success with lift in the top risk decile, calibrated probabilities for prioritization, and an offline to online check where the retention team tracked saved accounts. Include what you learned and what you would improve, like better point-in-time features, clearer monitoring, or a simpler model for easier adoption. Interviewers listen for ownership and clarity, so avoid jargon and be explicit about how your work changed a decision or process.

Data science interviews usually combine practical coding, statistics and machine learning concepts, plus business thinking and communication. You can expect a mix of quick screens, a technical round focused on SQL and Python, a case or take-home project, and behavioral questions about how you work. The good news is you do not need to know everything, but you do need a clear way to think. If you practice explaining your choices, tradeoffs, and results in plain language, you will stand out even when the questions get tough.

Common Interview Questions

Behavioral Questions (STAR Method)

STAR Method: Structure your answers using Situation, Task, Action, and Result to tell compelling stories about your experience.

Questions to Ask the Interviewer

Show your interest by asking thoughtful questions

•What does success look like for this role after 30, 60, and 90 days, and how will it be measured?
•How does your team decide whether a model should be shipped, and what metrics are treated as non-negotiable guardrails?
•What does the data stack look like today, and where do data quality or access issues slow the team down?
•How do you handle experimentation when randomization is hard, for example, network effects, small samples, or long feedback cycles?
•Can you share an example of a recent data science project that did not work out, and what the team learned from it?

Interview Preparation Tips

1
Practice answering data science interview questions out loud, and time yourself so your explanations stay clear and under two minutes.
2
Build a small set of stories you can reuse, one impact project, one failure, one conflict, and one ambiguous problem, and map each to the job requirements.
3
For technical prep, recreate common tasks from scratch, window functions in SQL, basic model evaluation, and point-in-time feature building, so you are not relying on memory alone.
4
When you get stuck, narrate your assumptions and ask clarifying questions, interviewers usually care more about your thinking than a fast guess.

data science Interview Questions: Complete Guide

Emily Thompson

Common Interview Questions

Behavioral Questions (STAR Method)

Questions to Ask the Interviewer

Interview Preparation Tips

Related guides

nurse Interview Questions: Complete Guide

angular Interview Questions: Complete Guide

business analyst Interview Questions: Complete Guide

html css Interview Questions: Complete Guide

data science Interview Questions: Complete Guide

Emily Thompson

Common Interview Questions

Q1How would you explain the difference between supervised and unsupervised learning, and when you would use each?

Q2Walk me through how you would evaluate a classification model.

Q3What is data leakage, and how do you prevent it?

Q4How would you design an A/B test, and what pitfalls would you watch for?

Q5Explain bias-variance tradeoff in practical terms.

Q6How do you handle missing data, and how do you choose an imputation strategy?

Q7Describe how you would approach a SQL question where you need the top N per group.

Q8How would you choose between logistic regression, random forest, and gradient boosting for a classification problem?

Q9How do you detect and handle outliers?

Q10Tell me about a data science project you are proud of, and how you measured success.

Behavioral Questions (STAR Method)

B1Tell me about a time you disagreed with a stakeholder about a model or metric.

B2Describe a time your analysis was wrong or incomplete. What did you do next?

B3Tell me about a time you had to explain a complex result to a non-technical audience.

Questions to Ask the Interviewer

Interview Preparation Tips

Related guides

nurse Interview Questions: Complete Guide

angular Interview Questions: Complete Guide

business analyst Interview Questions: Complete Guide

html css Interview Questions: Complete Guide