A data science interview is not a contest to name the most complicated model before anyone else. It tests whether you can turn a messy problem into a measurable one, choose a sensible approach, evaluate it correctly, and explain what happens when the model meets reality.
A simpler model with a clear reason often beats a fashionable model introduced with jazz hands.
A useful order for machine learning case questions
- 1
Define the decision
Who will use the prediction, what will they do with it, and what does a wrong answer cost?
- 2
Define the target and data
Explain labels, timing, availability, bias, and whether the data will exist at prediction time.
- 3
Choose a baseline
Start with a simple rule or model so improvement has a real comparison.
- 4
Choose evaluation
Match metrics to the decision and include offline, experiment, and monitoring plans.
- 5
Plan for failure
Discuss drift, feedback loops, unusual cases, and when a human should remain involved.
Model and evaluation questions
QUESTION 01
When would you choose a simpler model?
Why they ask: This tests whether complexity serves the problem rather than your ego.
- Compare performance against the baseline
- Include latency, explainability, maintenance, and data size
- State what benefit would justify more complexity
Likely follow-up: How much performance would you trade for explainability?
QUESTION 02
How would you detect data leakage?
Why they ask: This checks whether evaluation reflects information available at prediction time.
- Review when every feature is created
- Keep training and test work separate
- Investigate suspiciously strong results
Likely follow-up: Can preprocessing cause leakage?
QUESTION 03
Which metric would you use for an imbalanced classification problem?
Why they ask: This tests whether the metric matches the real error costs.
- Ask what false positives and false negatives cost
- Discuss precision, recall, thresholds, and calibration as needed
- Avoid choosing one metric without the decision context
Likely follow-up: Who chooses the threshold?
What interviewers may ask about your project
| Part | Question to prepare | Weak answer |
|---|---|---|
| Problem | Why did this need machine learning? | Because we had a lot of data |
| Data | How were labels created and checked? | The dataset was already clean |
| Model | What baseline and alternatives did you compare? | This model is state of the art |
| Evaluation | Why did the metric fit the decision? | It gave the highest score |
| Launch | What changed after deployment? | The model was successfully deployed |
| Limits | Where should the model not be trusted? | We did not find any major limitations |
A failed model can make a strong interview story
Useful failure
“Our churn model performed well offline but produced too many alerts for the customer team to use. I had optimised recall without checking the team’s weekly capacity. We worked backward from the number of customers they could contact, adjusted the decision threshold, and added account value to the ranking. The smaller list produced more completed calls and a higher save rate.”
The answer owns the mistake, connects the metric to operations, and shows a better decision.
Statistics topics to explain in plain language
- Sampling: Who is represented, who is missing, and why that matters.
- Experiments: Randomisation, power, guardrails, and what can break the result.
- Uncertainty: What an estimate can and cannot support.
- Correlation and cause: Which other explanations remain possible.
- Evaluation: Why one metric is useful for this decision and not every decision.
Data science interview checklist
- Every model answer starts with the decision
- A simple baseline is included
- Metrics match real error costs
- One project failure story is ready
- Data leakage and bias are considered
- You can explain the model to a non-specialist
Sources and further reading
These links support the advice above and give you somewhere useful to continue reading.
A model answer needs follow-up testing too
Practise defending your data science choices
Work through model, metric, project, and business questions in a realistic interview.
Start a data science interview