ML Interview Questions

Introduction

Machine learning interviews evaluate your ability to reason about algorithms, design systems, and make engineering tradeoffs. Interviewers look for structured thinking, mathematical intuition, and practical experience. This guide covers the most common questions organized by topic, with detailed answers that demonstrate the depth of understanding expected at top tech companies.

Supervised Learning Questions

Q1: Explain the bias-variance tradeoff with a concrete example.

Prediction	price = $200K + $50 × bedrooms
Reality	Prices depend on location, size, condition, market, etc.
Error	Consistently underpredicts mansions, overpredicts shacks
Training error	High \| Test error: High
On new data	Wild predictions between training points
Training error	~0 \| Test error: Very high
Training error	Moderate \| Test error: Moderate

Q2: When would you use Random Forest vs Gradient Boosting?

Random Forest

- Trains trees independently (parallelizable)

- Less prone to overfitting (bagging reduces variance)

- Less hyperparameter tuning needed

- Better when: Limited tuning time, noisy data, need robust baseline

Gradient Boosting (XGBoost/LightGBM)

- Trains trees sequentially (each corrects previous errors)

- Can overfit if not regularized (reduce bias)

- Typically higher accuracy with proper tuning

- Better when: You can invest time in tuning, structured/tabular data

Key difference

RF: Many deep trees averaged → reduce VARIANCE

GB: Many shallow trees summed → reduce BIAS

Q3: How do you handle class imbalance?

Scenario	Fraud detection (0.1% positive, 99.9% negative)
- Oversampling minority class (SMOTE	synthesize new minority samples)
- Combination	SMOTE + undersampling majority
- Class weights	Give minority class higher loss weight
- Threshold adjustment	Instead of 0.5, use lower threshold for positive
- USE	Precision-Recall curve, F1, AUC-ROC

Q4: Explain cross-validation and when stratified K-fold matters.

K-Fold Cross-Validation

1. Split data into K equal folds (typically K=5 or 10)

2. For each fold i:

- Train on all folds except i

- Evaluate on fold i

3. Average performance across all K evaluations

Fold 1: [Test] [Train] [Train] [Train] [Train]

Fold 2: [Train] [Test] [Train] [Train] [Train]

Fold 3: [Train] [Train] [Test] [Train] [Train]

Fold 4: [Train] [Train] [Train] [Test] [Train]

Fold 5: [Train] [Train] [Train] [Train] [Test]

Stratified K-Fold

Maintains class distribution in each fold

Critical when: Classes are imbalanced

Without stratification: Some folds might have 0 minority samples!

Deep Learning Questions

Q5: What happens if learning rate is too high or too low?

Symptom: Loss	NaN or spikes repeatedly
Symptom	Loss barely changes over many epochs
- LR range test	Increase LR exponentially, plot loss vs LR
- Typical ranges	1e-4 to 1e-2 for Adam, 1e-2 to 0.1 for SGD

Q6: Explain attention mechanism in transformers.

Core idea	Let each token "look at" all other tokens to gather context
Input	"The cat sat on the mat"
Query (what am I looking for?)	derived from "sat" embedding
Key (what do I contain?)	derived from each token's embedding
Value (what information do I provide?)	derived from each token
Attention scores	Q_sat · K_the, Q_sat · K_cat, Q_sat · K_sat, ...
After softmax	[0.05, 0.35, 0.10, 0.15, 0.05, 0.30]
Output	Weighted sum of Values using these attention weights
Result	"sat" representation enriched with info about WHO sat WHERE

System Design Questions

Q7: Design a content recommendation system for a news app.

- Users	10M daily active
- Content	10K new articles/day
- Latency	<100ms per request
- Optimize	Click-through rate + reading time
- Collaborative filtering	users like you read X
- Content-based	similar to articles you've read
- Trending	popular in your region/interests
Ranking (score 1000	select top 20):
- Features	user history, article embeddings, freshness,
- Model	Deep neural network predicting P(click) and P(read)
- Diversity	Don't show all articles from same topic
- Freshness	Boost recently published content
- Quality	Penalize clickbait (low read-time/click ratio)

Practical Coding Questions

Example

Common coding tasks in ML interviews: 1. Implement k-means clustering from scratch 2. Code logistic regression with gradient descent 3. Write data preprocessing pipeline (handle nulls, encode categoricals) 4. Implement cross-validation loop 5. Calculate precision, recall, F1 from confusion matrix 6. Code a simple neural network forward pass Tip: Practice coding these without libraries first, then show you know the library shortcuts (sklearn, PyTorch)

Summary

Machine learning interviews test breadth (knowing many algorithms), depth (understanding mathematical foundations), and practical wisdom (knowing which approach to use when). Structure your answers clearly: state the concept, give the formula or algorithm, provide intuition for why it works, discuss tradeoffs, and give a practical example. The best candidates demonstrate they can reason through novel problems using fundamental principles, not just recall memorized definitions.

Introduction

Supervised Learning Questions

Q1: Explain the bias-variance tradeoff with a concrete example.

Prediction	price = $200K + $50 × bedrooms
Reality	Prices depend on location, size, condition, market, etc.
Error	Consistently underpredicts mansions, overpredicts shacks
Training error	High \| Test error: High
On new data	Wild predictions between training points
Training error	~0 \| Test error: Very high
Training error	Moderate \| Test error: Moderate

Q2: When would you use Random Forest vs Gradient Boosting?

Random Forest

- Trains trees independently (parallelizable)

- Less prone to overfitting (bagging reduces variance)

- Less hyperparameter tuning needed

- Better when: Limited tuning time, noisy data, need robust baseline

Gradient Boosting (XGBoost/LightGBM)

- Trains trees sequentially (each corrects previous errors)

- Can overfit if not regularized (reduce bias)

- Typically higher accuracy with proper tuning

- Better when: You can invest time in tuning, structured/tabular data

Key difference

RF: Many deep trees averaged → reduce VARIANCE

GB: Many shallow trees summed → reduce BIAS

Q3: How do you handle class imbalance?

Scenario	Fraud detection (0.1% positive, 99.9% negative)
- Oversampling minority class (SMOTE	synthesize new minority samples)
- Combination	SMOTE + undersampling majority
- Class weights	Give minority class higher loss weight
- Threshold adjustment	Instead of 0.5, use lower threshold for positive
- USE	Precision-Recall curve, F1, AUC-ROC

Q4: Explain cross-validation and when stratified K-fold matters.

K-Fold Cross-Validation

1. Split data into K equal folds (typically K=5 or 10)

2. For each fold i:

- Train on all folds except i

- Evaluate on fold i

3. Average performance across all K evaluations

Fold 1: [Test] [Train] [Train] [Train] [Train]

Fold 2: [Train] [Test] [Train] [Train] [Train]

Fold 3: [Train] [Train] [Test] [Train] [Train]

Fold 4: [Train] [Train] [Train] [Test] [Train]

Fold 5: [Train] [Train] [Train] [Train] [Test]

Stratified K-Fold

Maintains class distribution in each fold

Critical when: Classes are imbalanced

Without stratification: Some folds might have 0 minority samples!

Deep Learning Questions

Q5: What happens if learning rate is too high or too low?

Symptom: Loss	NaN or spikes repeatedly
Symptom	Loss barely changes over many epochs
- LR range test	Increase LR exponentially, plot loss vs LR
- Typical ranges	1e-4 to 1e-2 for Adam, 1e-2 to 0.1 for SGD

Q6: Explain attention mechanism in transformers.

Core idea	Let each token "look at" all other tokens to gather context
Input	"The cat sat on the mat"
Query (what am I looking for?)	derived from "sat" embedding
Key (what do I contain?)	derived from each token's embedding
Value (what information do I provide?)	derived from each token
Attention scores	Q_sat · K_the, Q_sat · K_cat, Q_sat · K_sat, ...
After softmax	[0.05, 0.35, 0.10, 0.15, 0.05, 0.30]
Output	Weighted sum of Values using these attention weights
Result	"sat" representation enriched with info about WHO sat WHERE

System Design Questions

Q7: Design a content recommendation system for a news app.

- Users	10M daily active
- Content	10K new articles/day
- Latency	<100ms per request
- Optimize	Click-through rate + reading time
- Collaborative filtering	users like you read X
- Content-based	similar to articles you've read
- Trending	popular in your region/interests
Ranking (score 1000	select top 20):
- Features	user history, article embeddings, freshness,
- Model	Deep neural network predicting P(click) and P(read)
- Diversity	Don't show all articles from same topic
- Freshness	Boost recently published content
- Quality	Penalize clickbait (low read-time/click ratio)

Practical Coding Questions

Example

ML Interview Questions

Introduction

Supervised Learning Questions

Q1: Explain the bias-variance tradeoff with a concrete example.

Q2: When would you use Random Forest vs Gradient Boosting?

Random Forest

Gradient Boosting (XGBoost/LightGBM)

Key difference

Q3: How do you handle class imbalance?

Q4: Explain cross-validation and when stratified K-fold matters.

K-Fold Cross-Validation

Stratified K-Fold

Deep Learning Questions

Q5: What happens if learning rate is too high or too low?

Q6: Explain attention mechanism in transformers.

System Design Questions

Q7: Design a content recommendation system for a news app.

Practical Coding Questions

Summary

Continue learning this concept

ML Interview Questions

Introduction

Supervised Learning Questions

Q1: Explain the bias-variance tradeoff with a concrete example.

Q2: When would you use Random Forest vs Gradient Boosting?

Random Forest

Gradient Boosting (XGBoost/LightGBM)

Key difference

Q3: How do you handle class imbalance?

Q4: Explain cross-validation and when stratified K-fold matters.

K-Fold Cross-Validation

Stratified K-Fold

Deep Learning Questions

Q5: What happens if learning rate is too high or too low?

Q6: Explain attention mechanism in transformers.

System Design Questions

Q7: Design a content recommendation system for a news app.

Practical Coding Questions

Summary

Continue learning this concept