Data scientist interview questions: explain the decision

Use a question-to-action answer flow

Name the user, operational or policy question, the consequence of being wrong and what a useful answer would enable.

Explain data quality, assumptions, comparison, validation and the collaborators who improved the work.

Share the recommendation, uncertainty, safeguard and follow-up instead of presenting a model as an automatic answer.

This structure makes technical work legible without simplifying it beyond recognition. It also keeps credit accurate. A reliable data product may involve domain specialists, engineers, analysts, researchers, security colleagues and decision-makers. State your contribution clearly and give others their due.

Build a five-story data science bank

For each example, note the original question, the data you could and could not use, the test or method, your role, the result and one limitation. A limitation might be sample size, changing behaviour, missing labels, a proxy measure or a decision that needed monitoring. Naming it makes the answer more credible, not less capable.

• A framing story: you challenged a vague request, clarified the decision and agreed what success should mean.
• A data-quality story: you found a gap, bias, definition problem or unreliable source and changed the work responsibly.
• A method story: you compared approaches, selected a proportionate technique and explained why it fit the evidence.
• An ethics story: you identified a privacy, fairness, consent or misuse risk and involved the right people before proceeding.
• An adoption story: you communicated a finding, limitation or monitoring need so a team could use it without false certainty.

Handle a modelling case with disciplined curiosity

Ask how the output will be used, what error matters, what constraints apply and what would count as useful evidence.

Check provenance, completeness, timeliness, representativeness and whether a target or proxy could create harm.

Describe a baseline, validation plan and how you would compare alternatives before declaring a winner.

Name owners, monitoring, user feedback, drift or escalation conditions. Deployment is a handover, not the end of responsibility.

Original practice questions

Start with the decision and audience. Explain how you understood the data, what you tested, how you communicated the result and what happened next. Be precise about whether the work informed a choice, changed a process or simply reduced uncertainty.

Discuss the intended use before the metric. Explain data checks, a baseline, relevant measures, validation, ethical and privacy review, operational ownership and monitoring. A model can perform well on a test set and still be unsuitable for a real decision.

Name the issue without blaming its source. Explain how you investigated its likely effect, communicated it, chose a proportionate correction or limitation and documented the follow-up. Avoid claiming that cleaned data is automatically unbiased.

The UK Government Digital and Data Profession data scientist framework , last updated 29 August 2025, covers statistical practice, data engineering, ethics and privacy, delivery, programming and communication. Use it to organise your evidence, not to predict an employer's questions.

Turn the role requirements into a decision evidence map.

Match each requirement to a decision, evidence, method, collaborator, limitation and outcome that you can explain truthfully.