Question 1

How do you decide between a simple model like logistic regression and a more complex model like gradient boosting?

Accepted Answer

I start with the simple model as a baseline since it's interpretable and fast to iterate on, then check if the complex model's accuracy gain justifies the added latency and maintenance cost. For something like fraud detection where marginal accuracy matters a lot, I'll push toward XGBoost; for a regulated credit decision, interpretability often wins even at some accuracy cost.

Question 2

Walk me through how you'd handle a dataset with significant class imbalance.

Accepted Answer

I avoid relying on accuracy and use metrics like precision, recall, and PR-AUC instead, then address the imbalance with class weighting or techniques like SMOTE rather than just downsampling blindly. I also check that my train and test splits preserve the imbalance ratio so I'm evaluating on a realistic distribution.

Question 3

How do you validate that a model isn't overfitting before deployment?

Accepted Answer

I use k-fold cross-validation, typically 5-fold, and check that performance is stable across folds rather than trusting a single train-test split. I also hold out a true final test set that I only touch once, at the very end, to get an honest read on generalization.

Question 4

Describe how you would explain a model's prediction to a non-technical stakeholder.

Accepted Answer

I use SHAP values or feature importance to identify the top three or four drivers of a specific prediction, then translate them into plain business language, like 'this customer's churn risk is high mainly due to a drop in usage over the last 30 days.' I avoid showing raw coefficients or technical jargon in that conversation.

Question 5

How do you approach feature engineering for a time series forecasting problem?

Accepted Answer

I create lag features and rolling averages at intervals that match the business cycle, like weekly and monthly, and I'm careful to only use information that would have actually been available at prediction time to avoid leakage. I also test for seasonality and trend before choosing between something like ARIMA or a tree-based model with time features.

Question 6

What's your process when a production model's performance starts degrading over time?

Accepted Answer

I check for data drift first, comparing the distribution of current input features against the training distribution using something like population stability index. If drift is confirmed, I retrain on more recent data rather than assuming the original model architecture is the problem.

Question 7

How do you handle missing data in a dataset before modeling?

Accepted Answer

I first check whether the missingness is random or systematic, since that changes the right approach, then use imputation like median or model-based methods for random gaps, or a missing indicator flag if the absence itself is informative. I never just drop rows without understanding why the data is missing first.

Question 8

Describe how you would set up an A/B test to evaluate a new recommendation model.

Accepted Answer

I define the primary metric upfront, like click-through rate, calculate the required sample size for statistical power before starting, and randomize at the user level to avoid contamination between groups. I run it for a full business cycle, usually at least one to two weeks, to account for day-of-week effects.

Question 9

How do you decide what to log and monitor for a model in production?

Accepted Answer

I track prediction distribution, input feature distributions, and the actual business metric like conversion rate, not just model accuracy, since ground truth labels are often delayed. I set alert thresholds so I know within a day if something looks off rather than finding out weeks later from a stakeholder.

Question 10

Walk me through how you'd approach a request from a stakeholder who wants '99% accuracy' on a hard prediction problem.

Accepted Answer

I'd ask what decision the number actually drives, then show them the realistic performance ceiling given the data and explain tradeoffs like precision versus recall for their use case. Setting an arbitrary accuracy target without understanding the base rate or business cost of errors usually leads to a model optimized for the wrong thing.

Data Scientist interview questions and sample answers

Job-specific questions

1. How do you decide between a simple model like logistic regression and a more complex model like gradient boosting?

2. Walk me through how you'd handle a dataset with significant class imbalance.

3. How do you validate that a model isn't overfitting before deployment?

4. Describe how you would explain a model's prediction to a non-technical stakeholder.

5. How do you approach feature engineering for a time series forecasting problem?

6. What's your process when a production model's performance starts degrading over time?

7. How do you handle missing data in a dataset before modeling?

8. Describe how you would set up an A/B test to evaluate a new recommendation model.

9. How do you decide what to log and monitor for a model in production?

10. Walk me through how you'd approach a request from a stakeholder who wants '99% accuracy' on a hard prediction problem.

Behavioral questions

11. Tell me about a result you are proud of.

12. Describe a time you improved how something was done.

13. Tell me about a time you had to deliver under pressure.

14. Give an example of working with a difficult colleague or customer.

15. What is something you learned from a mistake?

Before the interview

Interview questions for related jobs