Data scientist interview questions

Six questions to practice out loud

Written by AI from data scientist job postings. Each one is the kind of question a data scientist interview asks, and each tests one of the five skills you are scored on.

1. QUESTION 1 · ROLE RELEVANCE

   Tell me about yourself through two analyses that changed a decision someone made, in under two minutes.

   They are listening for: It picks two projects and gives each the question, the data, the method in a line and the decision it changed, with a number.

   What a strong answer includes

   It picks two projects and gives each the question, the data, the method in a line and the decision it changed, with a number. It ends on why this team. Listing tools and languages, or walking the resume in order, scores low.

2. QUESTION 2 · COMMUNICATION & PRESENCE

   Explain a model you built to me as if I run the program it affects and have never trained a model. What does it do, how often is it wrong, and what should I not use it for?

   They are listening for: It says what the model predicts and which decision it feeds, in one plain sentence.

   What a strong answer includes

   It says what the model predicts and which decision it feeds, in one plain sentence. The error rate comes in terms the listener can picture, such as how many flagged cases turn out fine, along with one situation where the model should not be trusted. A weak answer names the algorithm and an accuracy score and stops.

3. QUESTION 3 · JUDGMENT & PROBLEM-SOLVING

   Tell me about a model that looked good in testing and disappointed once people used it. How did you find out, what was actually wrong, and what did you change?

   They are listening for: It names the metric that looked good, the signal that showed the problem and who raised it, and the cause, such as drift, leakage or a population the training data mis...

   What a strong answer includes

   It names the metric that looked good, the signal that showed the problem and who raised it, and the cause, such as drift, leakage or a population the training data missed. Then it says what changed in the model and in how you validate now. Blaming the data or the users, with nothing changed, is the weak version.

4. QUESTION 4 · JUDGMENT & PROBLEM-SOLVING

   You are handed a dataset for an analysis due Friday and notice one group of people is badly underrepresented in it. Walk me through what you check, what you tell the person who asked, and whether you go ahead.

   They are listening for: It checks how the data was collected before touching a model, sizes the gap with a number, and says what the analysis can and cannot claim as a result.

   What a strong answer includes

   It checks how the data was collected before touching a model, sizes the gap with a number, and says what the analysis can and cannot claim as a result. The requester hears early, with a choice: a narrower question, a stated caveat or a later date. A weak answer ignores the gap or refuses to deliver anything.

5. QUESTION 5 · SPECIFICITY & EVIDENCE

   Tell me about a vague request you turned into an analysis someone acted on. What was the original ask, what question did you actually answer, and what changed because of it?

   They are listening for: It quotes the original request, then the sharper question you agreed on and who agreed to it.

   What a strong answer includes

   It quotes the original request, then the sharper question you agreed on and who agreed to it. The data and method get a line each; the decision or number that moved afterward gets the most time. A weak answer describes the analysis in detail and never says what anyone did with it.

6. QUESTION 6 · ANSWER STRUCTURE

   Tell me about a recommendation from your analysis that a senior person did not want to hear. What did you show them, what did they push back on, and what did they decide?

   They are listening for: It states the finding, why it was unwelcome, the one number or chart you led with, and the objection in the other person's words.

   What a strong answer includes

   It states the finding, why it was unwelcome, the one number or chart you led with, and the objection in the other person's words. It ends on the decision, even one that went against you, and what you now do differently when you present. Making the other person the villain scores low.