Machine Learning Engineer interview questions
100 real questions with model answers and explanations for Junior candidates.
See a Machine Learning Engineer resume example →Practice with flashcards
Spaced repetition · Hunter Pass
Questions
The difference is whether the training data includes known target labels.
- Supervised learning maps inputs to provided targets.
- Unsupervised learning finds structure in unlabeled data.
- Classification and regression are supervised, while clustering is unsupervised.
Why interviewers ask this: The interviewer checks whether you can distinguish learning from labeled targets from discovering patterns without them.
Classification predicts categories, while regression predicts numerical values.
- A spam detector is a classification model.
- A house price predictor is a regression model.
- Both are supervised tasks when examples include target values.
Why interviewers ask this: A strong answer identifies the type of target produced by each task.
Features are model inputs, and the target is the value the model learns to predict.
- Features can be measurements, categories, or derived values.
- The target is available during supervised training.
- Feature quality strongly affects what patterns a model can learn.
Why interviewers ask this: The interviewer is evaluating whether you can correctly define the inputs and desired output of a supervised problem.
The three sets separate model fitting, model selection, and final evaluation.
- The training set is used to learn model parameters.
- The validation set is used to tune hyperparameters and compare models.
- The test set estimates performance on unseen data after all choices are fixed.
Why interviewers ask this: The interviewer checks whether you know how separate datasets prevent an overly optimistic evaluation.
Overfitting means learning the training data too closely, while underfitting means missing useful structure.
- An overfit model performs well on training data but poorly on new data.
- An underfit model performs poorly on both training and new data.
- Model complexity, regularization, and data volume affect both problems.
Why interviewers ask this: A strong answer relates model fit to both training performance and generalization.
The bias-variance tradeoff balances errors from models that are too simple and models that are too sensitive to training data.
- High bias often causes underfitting.
- High variance often causes overfitting.
- Good generalization requires an appropriate balance between them.
Why interviewers ask this: The interviewer wants to see that you connect model complexity with underfitting and overfitting.
L1 and L2 regularization penalize model weights in different ways.
- L1 adds a penalty based on absolute weight values.
- L2 adds a penalty based on squared weight values.
- L1 can produce zero weights, while L2 usually shrinks weights smoothly.
Why interviewers ask this: A strong answer explains both the penalty formulas and their different effects on weights.
Parameters are learned from data, while hyperparameters are chosen around the training process.
- Linear regression coefficients are model parameters.
- Learning rate and tree depth are hyperparameters.
- Hyperparameters are commonly selected using validation results.
Why interviewers ask this: The interviewer checks whether you distinguish fitted values from settings that control model training or structure.
A loss function measures how far model predictions are from desired outputs.
- Training aims to minimize the loss.
- Different tasks need different loss functions.
- The loss gives the optimization algorithm a numerical objective.
Why interviewers ask this: A strong answer connects prediction error, optimization, and the choice of task-specific objective.
Mean squared error is the average squared difference between predictions and true values.
- It is commonly used for regression.
- Squaring makes every contribution nonnegative.
- Large errors receive a stronger penalty than small errors.
Why interviewers ask this: The interviewer checks whether you understand why MSE emphasizes large regression errors.
Cross-entropy measures how well predicted class probabilities match the true classes.
- It is commonly used for classification.
- Confident incorrect predictions receive a large penalty.
- Lower cross-entropy means better probability assignments to the true labels.
Why interviewers ask this: A strong answer explains that cross-entropy evaluates probabilities rather than only final class labels.
Gradient descent repeatedly updates model parameters in the direction that reduces the loss.
- The gradient shows how the loss changes with each parameter.
- Parameters move in the opposite direction of the gradient.
- Repeated updates can approach a local loss minimum.
Why interviewers ask this: The interviewer is evaluating whether you understand how derivatives guide parameter updates.
The learning rate controls the size of each parameter update during optimization.
- A rate that is too small can make training slow.
- A rate that is too large can overshoot good parameter values.
- A suitable rate helps training converge reliably.
Why interviewers ask this: A strong answer connects the learning rate to both optimization speed and stability.
They differ in how many training examples are used for each parameter update.
- Batch gradient descent uses the full training set.
- Stochastic gradient descent uses one example at a time.
- Mini-batch gradient descent uses a small group and is common in practice.
Why interviewers ask this: The interviewer checks whether you understand how batch size changes update cost, frequency, and noise.
Feature scaling puts numerical features on comparable scales.
- It prevents large-valued features from dominating distance calculations.
- It can make gradient-based optimization converge faster.
- Trees usually need scaling less than k-nearest neighbors or linear models.
Why interviewers ask this: A strong answer identifies which model families are sensitive to feature scale.
Standardization centers values by their mean and spread, while normalization usually maps them to a fixed range.
- Standardization commonly produces mean zero and standard deviation one.
- Min-max normalization commonly maps values between zero and one.
- Both transformations must be fitted on training data only.
Why interviewers ask this: The interviewer checks whether you know the transformations and how to apply them without leakage.
Missing values can be removed, imputed, or represented explicitly.
- Drop rows or columns only when the data loss is acceptable.
- Impute with statistics such as the median or with model-based estimates.
- Add a missing-value indicator when absence may carry information.
Why interviewers ask this: A strong answer presents several methods and recognizes that the reason for missingness matters.
Categorical variables are converted into numerical representations that a model can process.
- One-hot encoding creates a binary column for each category.
- Ordinal encoding assigns ordered numeric values.
- The encoder must learn its categories from training data only.
Why interviewers ask this: The interviewer evaluates whether you can encode categories without inventing false relationships or leaking data.
Ordinal encoding suits categories with a meaningful order, while one-hot encoding suits unordered categories.
- Sizes such as small, medium, and large have a natural rank.
- Colors usually have no natural order.
- Arbitrary numeric codes can create false relationships between categories.
Why interviewers ask this: A strong answer makes the encoding reflect the real structure of the feature.
Outliers are observations that differ substantially from most of the data.
- They may be valid rare cases or data errors.
- They can strongly affect means, MSE, and some fitted models.
- Their cause should be investigated before removal or transformation.
Why interviewers ask this: The interviewer checks whether you avoid deleting unusual values without understanding them.
Locked questions
- 21
What is feature engineering?
feature-engineering - 22
What is data leakage?
leakage - 23
What is class imbalance and how can it be handled?
imbalance - 24
What does accuracy measure in classification?
classification - 25
What is the difference between precision and recall?
evaluation - 26
What is the F1 score?
classification-metrics - 27
What does a confusion matrix show?
evaluation - 28
What does ROC-AUC measure?
evaluation - 29
What do the precision-recall curve and PR-AUC show?
evaluation - 30
How do MAE, RMSE, and R-squared evaluate regression models?
evaluationdecision-making - 31
What is a baseline model and why is it useful?
modeling - 32
How does k-fold cross-validation work?
cross-validationvalidation - 33
What is stratified cross-validation?
cross-validationvalidation - 34
How should time-series data be split for validation?
validation - 35
How does linear regression work, and what are its main assumptions?
regression - 36
How does logistic regression perform classification?
classificationregression - 37
How does a decision tree make predictions?
trees - 38
How does a random forest differ from a single decision tree?
ensemblestrees - 39
How does gradient boosting work?
boosting - 40
How does k-nearest neighbors make a prediction?
knn - 41
What is the main idea behind a support vector machine?
svm - 42
How does a naive Bayes classifier work?
bayesian - 43
How does k-means clustering work?
clustering - 44
What is principal component analysis used for?
components - 45
What happens during the forward pass of a neural network?
neural-nets - 46
What is backpropagation?
neural-networks - 47
Why do neural networks need activation functions?
neural-netsactivation - 48
What are an epoch, a batch, and an iteration?
batchiteration - 49
What is a scikit-learn Pipeline used for?
ci-cd - 50
What roles do NumPy, Pandas, scikit-learn, PyTorch, and TensorFlow play in ML?
pandasnumpypytorch - 51
You need to predict customer churn from tabular data. What would you build first?
churn - 52
Only 1% of transactions in a fraud dataset are fraudulent. How would you evaluate a classifier?
decision-makingtransactions - 53
A screening model must avoid missing people who may have a disease. Which metric would you prioritize?
monitoring - 54
Users complain that a spam filter hides legitimate emails. How would you adjust your evaluation?
- 55
Your classifier outputs probabilities, but the default threshold of 0.5 performs poorly. What would you do?
thresholding - 56
A model has 96% accuracy, but users say it misses most rare cases. How would you investigate?
- 57
A neural network is slightly more accurate than logistic regression, but it is much slower. Which would you choose?
neural-netsregression - 58
Training accuracy keeps improving while validation accuracy gets worse. How would you respond?
validation - 59
Both training and validation performance are poor. What would you check first?
validationperformance - 60
Validation accuracy suddenly jumps from 75% to 99% after adding one feature. What would you suspect?
validation - 61
You are predicting next week's demand from historical sales. How would you split the data?
- 62
The same customer can appear in many rows. How would you prevent an overly optimistic split?
locking - 63
A teammate scales all data before creating train and test sets. What problem does this cause?
train-test - 64
A loan model uses a field called final_collection_status. Would you keep it?
- 65
Production sends a category that your encoder never saw during training. How would you handle it?
- 66
Several numeric columns contain missing values. How would you handle them?
missing-dataschema - 67
A regression dataset contains a few extremely large values. What would you do before removing them?
- 68
Manual review shows that many training labels are wrong. How would you proceed?
- 69
How would you improve a classifier trained on highly imbalanced data?
imbalanced - 70
You have only a few thousand labeled examples. How would you approach modeling?
- 71
How would you use training, validation, and test sets during a project?
validation - 72
Cross-validation scores vary widely between folds. How would you investigate?
cross-validationvalidation - 73
A feature named record_id is the model's strongest feature. What would you do?
- 74
Your test score is good, but the test data comes from a different source than production. What would you do?
test-data - 75
An offline metric improves, but an A/B test shows no product benefit. How would you investigate?
ab-testingmonitoring - 76
A teammate cannot reproduce your best experiment. What would you provide?
experimentsdebugging - 77
The training data changes every week. How would you keep experiments comparable?
experiments - 78
What would you record for each model experiment?
experiments - 79
How would you turn a notebook prototype into a repeatable training workflow?
prototypes - 80
What would you test in a preprocessing pipeline?
ci-cd - 81
How would you expose a scikit-learn model through a basic prediction API?
api - 82
When would you choose batch predictions instead of an online API?
batchapi - 83
A prediction endpoint is too slow. How would you debug it?
endpoints - 84
Your model uses too much memory to fit on the target server. What would you try?
memory - 85
What input checks would you add to a model endpoint?
endpoints - 86
How would you release a new model without sending all traffic to it immediately?
- 87
What would make a model deployment easy to roll back?
deploymentrollback - 88
Which service metrics would you monitor for a prediction API?
monitoringapi - 89
How would you detect that production inputs have drifted from training data?
iac - 90
Ground-truth labels arrive 30 days after predictions. How would you monitor model quality?
monitoring - 91
When would you retrain a deployed model?
retrainingdeployment - 92
A user reports one obviously wrong prediction. How would you debug it?
- 93
Offline predictions differ from production predictions for the same record. What would you check?
- 94
An upstream team renames a feature column and predictions start failing. How would you handle the incident and prevent a repeat?
incidentsschema - 95
A product manager asks whether your model will be 90% accurate in production. How would you answer?
- 96
You are stuck debugging a training pipeline. When and how would you ask for help?
problem-solvingci-cd - 97
A reviewer disagrees with your model implementation. How would you handle it?
conflict - 98
A teammate shares a notebook with a promising result that nobody else can reproduce. What would you do?
- 99
A deadline is close, and the best model misses the latency requirement. What would you propose?
estimationlatency - 100
Tell me about an ML project that did not work as planned. What should a good answer include?
story