Machine learning engineer interview questions with concise answers
Machine learning engineer interviews test whether you can build a model that holds up outside the notebook. The strong answer names the metric, the baseline and the main risk, such as leakage or training-serving skew, and says how you would validate and monitor it. This guide gives twenty questions with short answers and follows the scikit-learn documentation and Google's Rules of Machine Learning.
How machine learning questions appear in a live interview
Expect fundamentals, a scenario such as choosing metrics for imbalanced data, a Python or SQL task, and a design question on taking a model to production. State your assumptions and the cost of each error type. The same habit helps in a technical phone screen and a live coding interview.
Junior machine learning engineer interview questions
1. What is the difference between supervised and unsupervised learning?
Supervised learning trains on labelled examples to predict a target, such as price or a class. Unsupervised learning finds structure without labels, such as clusters or lower-dimensional representations. Give one example of each and say which your last project used.
2. What is overfitting and how do you reduce it?
Overfitting is when a model memorises training data and performs worse on unseen data. Reduce it with more or better data, a simpler model, regularisation, early stopping and honest validation. Fitting and scoring on the same data gives misleadingly perfect results, as the scikit-learn cross-validation guide warns.
3. What is the bias-variance trade-off?
High bias means the model is too simple and misses patterns; high variance means it is too sensitive to the training sample. A more flexible model lowers bias but tends to raise variance. Use validation curves to find a sensible middle, and note that more data mostly helps variance.
4. What are training, validation and test sets for?
Train to fit parameters, validate to choose models and hyperparameters, and test once at the end for an unbiased estimate. If you tune against the test set, information leaks into the model and the estimate is too optimistic.
5. What is gradient descent?
An iterative method that moves parameters in the direction that reduces the loss, using the gradient. The learning rate sets the step size: too small is slow, too large can be unstable. In practice you use mini-batches and an optimiser such as SGD or Adam.
6. What is regularisation?
A penalty or constraint that discourages overly complex models, for example L1 or L2 penalties on weights, dropout or early stopping. It trades a little training fit for better generalisation. Strength is a hyperparameter you tune on validation data.
7. When would you use logistic regression instead of a neural network?
When you need a strong, fast, interpretable baseline on tabular data with limited samples. Start simple and add complexity only if it earns its keep. Google's Rules of ML say to keep the first model simple and get the infrastructure right.
Mid-level machine learning engineer interview questions
8. What is data leakage and how do you prevent it?
Leakage is using information at training time that will not exist at prediction time, which inflates scores. Split first, learn every transformation from the training data only, and wrap steps in a scikit-learn Pipeline. The common pitfalls page shows feature selection on all data producing spurious accuracy on random labels.
9. How does k-fold cross-validation work?
Split the training data into k folds, train on k minus 1, validate on the remaining one, and average the k scores. It uses data efficiently when samples are scarce. Keep a final test set aside, and use time-ordered splits for time series instead of random folds.
10. Accuracy is 98%. Why might that be misleading?
With imbalanced classes, a model that predicts the majority class can score high accuracy and be useless. Look at precision, recall, F1 and ROC AUC, defined in the scikit-learn model evaluation guide, and choose by the cost of each error type.
11. Explain precision and recall with an example.
Precision is the share of predicted positives that are correct: tp / (tp + fp). Recall is the share of real positives found: tp / (tp + fn). For fraud detection, high recall catches more fraud; high precision avoids blocking honest customers. The threshold moves you along the trade-off.
12. How do you handle imbalanced data?
Pick suitable metrics first, then consider class weights, resampling, collecting more minority examples and threshold tuning. Resample only inside the training folds to avoid leakage. Check that the metric reflects the business cost, not only a score.
13. How do you handle missing values and categorical features?
Understand why values are missing, then impute inside the pipeline using training data only, optionally with a missing-indicator. Encode categories with one-hot, ordinal or target-style encodings, taking care with high cardinality and leakage. Tree-based models tolerate some choices that linear models do not.
14. Walk me through a PyTorch training loop.
For each batch: forward pass to get predictions, compute the loss, call loss.backward() for gradients, optimizer.step() to update parameters, and optimizer.zero_grad() so gradients do not accumulate. Epochs, batch size and learning rate are hyperparameters, as in the PyTorch optimisation tutorial.
Senior machine learning engineer interview questions
15. What is training-serving skew and how do you detect it?
It is a difference between the data or code used in training and in production, so the model behaves worse live. Reuse the same feature code, log features at serving time, and compare distributions between training, holdout and live data. The Rules of ML suggest measuring skew at those points.
16. How do you monitor a model in production?
Track input data statistics, feature coverage, prediction distribution, latency and, when labels arrive, real quality metrics. Watch for silent failures such as stale tables. Define alert thresholds and an owner, and know how fast your model goes stale so you retrain on evidence, not a calendar.
17. How would you design an ML system from problem to production?
Clarify the goal and a baseline, possibly a heuristic. Define data, labels and metrics, build a simple model and a reproducible pipeline, evaluate offline, then release gradually with monitoring and a rollback. Revisit the problem framing if the metric and the business goal disagree.
18. How do you make experiments reproducible?
Version code, data and configuration, fix random seeds where it makes sense, log parameters and metrics for each run, and keep the environment pinned. Reproducibility lets you compare runs honestly and rebuild a model when something breaks.
19. How do you decide between a simple model and a complex one?
Compare against a strong baseline on the metric that matters, then weigh the gain against cost, latency, maintainability and interpretability. A small improvement may not justify a harder system. Say what gain would convince you.
20. Tell me about a model that failed in production.
Choose a real case. Describe the symptom, how you detected it, the root cause such as skew, drift or a bad label, the fix and the monitoring you added. It is judged on ownership and learning. The STAR method keeps it short.
A concise answer shape
For an ML question, say the goal and the metric, the baseline, the risk such as leakage or skew, and how you would validate and monitor. Example: for churn prediction, I would start with logistic regression, split by time to avoid leakage, judge it by recall at a fixed precision, and monitor feature drift after release.
Before any live assessment, ask the recruiter which tools and documentation are permitted. Knowing the material is useful; using unapproved tools during an assessment is a separate policy question. For a quiet rehearsal there is a free plan, with the same caveat. Related guides: Python interview questions, SQL interview questions, system design interview framework.
Sources
- scikit-learn: cross-validation and held-out data
- scikit-learn: common pitfalls, data leakage and Pipeline
- scikit-learn: model evaluation, precision, recall, F1, ROC AUC
- PyTorch: optimizing model parameters, the training loop
- Google: Rules of Machine Learning
Checked 12 November 2026.
FAQ
What are common machine learning engineer interview questions?
Overfitting and validation, bias and variance, metrics for imbalanced data, data leakage, feature handling, a coding task in Python, and how to deploy and monitor a model. Senior rounds add system design for ML.
What is the difference between a data scientist and an ML engineer interview?
Data scientist interviews lean on statistics, experimentation and analysis. ML engineer interviews add software engineering, pipelines, deployment and monitoring. Both test model evaluation, so know your metrics.
Which metrics should I explain in an ML interview?
Accuracy and its limits, precision, recall, F1 and ROC AUC for classification, and an error measure such as mean absolute error for regression. Tie each to the cost of mistakes in the example you give.
Do I need to know PyTorch or TensorFlow?
Know the one in the job ad well enough to explain a training loop, loss, optimiser and evaluation. Scikit-learn knowledge is usually expected for classical models and pipelines.
Will there be a coding round?
Usually: Python, data manipulation or SQL, sometimes implementing a simple algorithm. Ask the recruiter about the format and which tools and documentation are allowed.