Machine Learning Interview Questions for Freshers
Practise explaining leakage, metrics, baselines and model errors.
Part of the Machine Learning track: modules, projects and first roles →
By the GullyMentor team · 5 min read · Published 2026-10-18 · Updated 2026-11-06 · Sources listed at the end. Not a promise about mentorship outcomes.
Fresher machine learning interviews often test whether you can reason about a project you built. Use these questions to rehearse that reasoning. They are practice examples, not predictions of a particular employer's interview. For each, an outline of a good answer is given, then what a reviewer tends to look for.
What is overfitting?
Answer outline: a model fits development examples closely but performs poorly on unseen cases. Show it with the gap between training and validation scores. Then ask whether the split represents real use. Remedies include more data, simpler models and regularisation.
Reviewer looks for: a concrete example from your own notebook, and the point that none of the remedies replaces a credible evaluation setup.
What is data leakage?
Answer outline: information enters training that would not be available at prediction time. A churn model using a cancellation date is an obvious example. Preprocessing fitted on the whole dataset before splitting is a subtler one.
Reviewer looks for: whether you can show where your own pipeline separates fitting from evaluation. Do not stop at the definition.
Why use a baseline?
Answer outline: a baseline tests whether the model improves on a simple alternative. Predicting the previous week's demand may be hard to beat. If a complex model does worse, say you would investigate the data or keep the simpler approach.
Reviewer looks for: willingness to accept that a simple method can win.
Which metric would you choose?
Answer outline: begin with the decision and the error costs. For rare-event classification, compare precision and recall rather than reporting accuracy alone. For a delivery estimate, an error measured in minutes is interpretable. Work through a small example: 20 of 100 customers leave, the model flags 25 and catches 15, so precision is 60% and recall is 75%.
Reviewer looks for: a stated denominator and a reason tied to the problem. Avoid listing metrics without explaining when they help.
How would you handle missing values?
Answer outline: investigate why values are missing and whether missingness itself carries information. Choose a treatment fitted using training data only. Document the effect, and test behaviour when a new record lacks the field.
Reviewer looks for: awareness that replacing every missing value with zero can create a misleading signal.
What would you do after a good test score?
Answer outline: inspect errors, relevant groups, reproducibility and operational constraints. Ask whether the evaluation period represents future use. Note that a score does not establish permissions, reliability or a monitoring plan.
Reviewer looks for: healthy doubt, not celebration.
How do you explain the difference between regression and classification?
Answer outline: regression predicts a number, classification predicts a category. Give one example of each from the same dataset.
Reviewer looks for: clean, short examples. The guide on regression, classification and clustering can help you rehearse.
What does a strong answer sound like?
A strong answer is short, specific and honest. For example: "My churn model reached 85% accuracy, but the no-churn baseline reached 80%, so I looked at recall. It was 75%, and I found most misses were new customers, so I would collect more data on them." It names a number, a comparison, a finding and a next step. Aim for that pattern in every reply.
How should you practise?
1. Pick one project from your portfolio. 2. Answer each question above using that project's real details. 3. Record yourself for two minutes per answer and listen for vague phrases such as "it worked well". 4. Replace each vague phrase with a number, a table or a file name from your work.
Project rehearsal
Explain the question, target, available inputs, split, baseline and result in two minutes. Then discuss the largest error and what you changed after finding it.
If you have not deployed a model, say so. A clear notebook investigation is credible evidence of learning. Inventing production experience weakens it, and follow-up questions will expose it.
Common mistakes
- Memorising definitions without an example.
- Naming algorithms you cannot explain.
- Claiming a metric you cannot compute by hand on a small example.
- Hiding a failed experiment instead of describing what it taught you.
How do you check you are ready?
Ask a friend to interrupt your answer with "why?" three times. If your answer still holds up, you understand it. If you fall back to a definition, go back to your notebook.
What to build next
Strengthen the weak answers by finishing a project from machine learning project ideas and writing an evaluation note as in how to evaluate a machine learning model. See the Machine Learning track for the wider learning path.
Sources & further reading
Examples and practice recommendations are educational illustrations. Tool documentation and programme inclusions should be checked for your intended use.