1. 1. What's the difference between supervised and unsupervised learning?

   Supervised learning uses labeled data to predict outcomes, unsupervised finds patterns in unlabeled data. Semi-supervised combines both approaches.

2. 2. Explain overfitting and how to prevent it.

   Model memorizes training data instead of learning patterns. Prevent with regularization, cross-validation, early stopping, and more training data.

3. 3. What is bias-variance tradeoff?

   Bias is error from oversimplification, variance from sensitivity to small data changes. Total error = bias² + variance + irreducible error.

4. 4. When would you use classification vs regression?

   Classification for categorical outcomes (spam/not spam), regression for continuous values (house prices). Some problems can be framed either way.

5. 5. How do you handle imbalanced datasets?

   First, check if imbalance matters for your use case. Then consider resampling (SMOTE, undersampling), class weights, or different evaluation metrics like F1-score.

6. 6. Explain cross-validation and its types.

   Technique to assess model generalization. K-fold splits data into K parts, stratified maintains class distribution, time-series uses temporal splits.

7. 7. What's the difference between bagging and boosting?

   Bagging trains models in parallel on different data subsets (Random Forest), boosting trains sequentially, correcting previous mistakes (XGBoost).

8. 8. How would you explain a complex ML model to a non-technical stakeholder?

   Use analogies, focus on business impact, explain confidence levels, and demonstrate with simple examples. Avoid jargon and technical details.

9. 9. How do you design an ML system for real-time fraud detection?

   Consider latency requirements, feature engineering pipeline, model updating strategy, false positive handling, and gradual rollout with human oversight.

10. 10. What are the challenges of deploying ML models in production?

    Data drift, model decay, scalability, monitoring, versioning, A/B testing, rollback strategies, and ensuring consistent preprocessing.

11. 11. How would you handle concept drift?

    Monitor model performance metrics, detect drift using statistical tests, implement automated retraining pipelines, and use online learning when appropriate.

12. 12. Design an ML system architecture for a recommendation engine.

    Batch training for collaborative filtering, real-time feature serving, A/B testing framework, cold start handling, and scalable inference infrastructure.

13. 13. Compare decision trees vs random forests vs gradient boosting.

    Decision trees are interpretable but overfit. Random forests reduce variance through bagging. Gradient boosting sequentially corrects errors for higher accuracy.

14. 14. When would you use SVM over logistic regression?

    SVM for high-dimensional data, non-linear relationships (with kernels), and when you want maximum margin classification. Logistic regression for interpretability and probability estimates.

15. 15. Explain k-means clustering and its limitations.

    Partitions data into k spherical clusters. Limitations: assumes spherical clusters, sensitive to initialization, requires choosing k, affected by outliers.

16. 16. What's the difference between L1 and L2 regularization?

    L1 (Lasso) promotes sparsity and feature selection. L2 (Ridge) shrinks coefficients uniformly. Elastic Net combines both approaches.

17. 17. How does Principal Component Analysis (PCA) work?

    Finds orthogonal directions of maximum variance in data. Used for dimensionality reduction while preserving most information. Linear transformation based on eigenvectors.

18. 18. Explain gradient descent and its variants.

    Batch GD uses all data, SGD uses one sample, Mini-batch balances both. Adam adds momentum and adaptive learning rates for better convergence.

19. 19. What causes vanishing gradient problem and how to solve it?

    Gradients become very small in deep networks. Solutions: ReLU activation, skip connections, batch normalization, proper weight initialization.

20. 20. Compare CNNs, RNNs, and Transformers.

    CNNs for spatial data (images), RNNs for sequential data with memory, Transformers for parallel processing of sequences with attention mechanisms.

21. 21. What is transfer learning and when to use it?

    Using pre-trained models as starting point. Useful with limited data, similar domains, or when computational resources are constrained.

22. 22. Explain attention mechanism in neural networks.

    Allows models to focus on relevant parts of input. Computes weighted combinations based on query-key similarity, enabling better long-range dependencies.

23. 23. How do you evaluate a classification model beyond accuracy?

    Precision, recall, F1-score, ROC-AUC, PR curves. Consider business impact, class imbalance, and cost of false positives vs false negatives.

24. 24. What is A/B testing for ML models?

    Comparing model performance on real users. Split traffic between models, measure business metrics, control for confounding factors, ensure statistical significance.

25. 25. How do you handle missing data?

    Understand missingness pattern (MCAR, MAR, MNAR). Options: deletion, imputation (mean, median, KNN, model-based), or treat as separate category.

26. 26. What is feature scaling and when is it needed?

    Standardization (z-score) or normalization (min-max). Required for distance-based algorithms, gradient descent, neural networks. Not needed for tree-based methods.

27. 27. How do you detect and handle outliers?

    Statistical methods (IQR, z-score), visualization, or ML-based (Isolation Forest). Handle by removal, transformation, or robust algorithms.

28. 28. What is feature engineering and why is it important?

    Creating relevant features from raw data. Includes transformation, combination, selection. Often more impactful than choosing algorithms for model performance.

29. 29. How do you handle categorical variables?

    One-hot encoding for nominal, ordinal encoding for ordered categories, target encoding for high cardinality, embedding for deep learning models.

30. 30. What is data leakage and how to prevent it?

    Information from future or target variable in features. Prevent through proper train/test splits, understanding data generation process, and careful feature engineering.

31. 31. How do you monitor ML models in production?

    Track performance metrics, data quality, feature distributions, prediction latency, and business KPIs. Set up alerts for significant degradation.

32. 32. What is model versioning and why is it important?

    Track different model versions, training data, hyperparameters, and performance. Enables rollback, comparison, and reproducibility. Use tools like MLflow or DVC.

33. 33. How would you scale model inference for high traffic?

    Model optimization (quantization, pruning), caching, load balancing, batch processing, and distributed serving. Consider edge deployment for latency-sensitive applications.

34. 34. What are the key components of an ML pipeline?

    Data ingestion, preprocessing, feature engineering, model training, validation, deployment, monitoring, and retraining. Each component should be versioned and monitored.

35. 35. How do you ensure reproducibility in ML experiments?

    Version control code and data, set random seeds, document environment (Docker/Conda), track experiments, use configuration files, and automate workflows.

The PROBLEM Framework

Use this systematic approach for any ML question:

1. P - Problem Understanding: "Let me clarify the business objective and constraints..."
2. R - Requirements: "What are the accuracy, latency, and interpretability requirements?"
3. O - Options: "Here are three approaches we could consider..."
4. B - Best Approach: "Given the constraints, I'd recommend..."
5. L - Limitations: "The main risks and limitations are..."
6. E - Evaluation: "We'd measure success using these metrics..."
7. M - Monitoring: "In production, we'd monitor for..."

Common ML Interview Mistakes