Master Q&A Bank

---

Level 1: Foundations (Entry-Level / Junior)

Machine Learning Basics

1. What is Machine Learning?

ML is a subset of AI where systems learn patterns from data to make predictions or decisions without being explicitly programmed.

2. What's the difference between Supervised, Unsupervised, and Reinforcement Learning?

Supervised: Learns from labeled data (X→y). Unsupervised: Finds structure in unlabeled data. Reinforcement: Learns from environment feedback (rewards).

3. What is the Bias-Variance Trade-off?

Bias = error from simplifying assumptions (underfitting). Variance = error from sensitivity to noise (overfitting). Goal: minimize total error.

4. What is Overfitting? How do you prevent it?

Model learns noise in training data. Prevent with: regularization, more data, cross-validation, early stopping, dropout.

5. What is Underfitting?

Model is too simple to capture patterns. Fix with: more features, less regularization, more complex model.

6. Why do we split data into Train/Validation/Test?

Train: learn patterns. Validation: tune hyperparameters. Test: final unbiased evaluation.

7. What is Cross-Validation?

Split data into k folds, train on k-1, validate on 1, repeat k times. Gives robust performance estimate.

8. What is a Hyperparameter vs a Parameter?

Parameter: learned from data (weights). Hyperparameter: set before training (learning rate, k in KNN).

9. What is Feature Scaling? Why is it important?

Normalize features to similar ranges. Important for gradient-based algorithms and distance-based methods.

10. What is One-Hot Encoding?

Convert categorical variables to binary columns. One column per category.

Regression

11. What is Linear Regression?

Predicts continuous output as a linear combination of features: $y = w^Tx + b$.

12. What are the assumptions of Linear Regression?

Linearity, independence, homoscedasticity (constant variance), normality of errors.

13. What is R-squared?

Proportion of variance explained: $R^2 = 1 - \frac{SS_{res}}{SS_{tot}}$. Higher is better (max 1).

14. What is the difference between MAE and MSE?

MAE: Mean Absolute Error. Robust to outliers. MSE: Mean Squared Error. Penalizes large errors more.

15. What is Polynomial Regression?

Fits a polynomial curve by adding polynomial features ($x^2, x^3$, etc.) to linear regression.

Classification

16. What is Logistic Regression?

Classification algorithm using sigmoid function to output probabilities. Decision boundary is linear.

17. Why is it called "Regression" if it's for classification?

It "regresses" to a probability value, then thresholds for classification.

18. What is the Sigmoid function?

$\sigma(z) = \frac{1}{1 + e^{-z}}$. Maps any real number to (0, 1).

19. What is Cross-Entropy Loss?

$L = -[y\log(\hat{y}) + (1-y)\log(1-\hat{y})]$. Penalizes confident wrong predictions heavily.

20. What is a Confusion Matrix?

Table showing TP, TN, FP, FN. Used to calculate precision, recall, F1.

21. What is Precision?

$\frac{TP}{TP + FP}$. Of all predicted positive, how many were actually positive?

22. What is Recall (Sensitivity)?

$\frac{TP}{TP + FN}$. Of all actual positives, how many did we catch?

23. What is F1-Score?

Harmonic mean of Precision and Recall: $\frac{2 \cdot P \cdot R}{P + R}$. Balances both.

24. When do you use ROC-AUC vs PR-AUC?

ROC-AUC: Balanced datasets. PR-AUC: Imbalanced datasets (focuses on positive class).

25. What is a Decision Boundary?

The surface that separates different classes in feature space.

---

Level 2: Intermediate (Mid-Level / L4)

Model Comparison & Selection

26. L1 vs L2 Regularization?

L1 (Lasso): Adds $\sum|w|$. Creates sparse weights (feature selection). L2 (Ridge): Adds $\sum w^2$. Shrinks weights smoothly.

27. When to use Random Forest vs XGBoost?

RF: Robust baseline, less tuning. XGBoost: Higher accuracy, needs tuning, prone to overfit.

28. What is Bagging vs Boosting?

Bagging: Parallel independent models, reduces variance. Boosting: Sequential models correcting errors, reduces bias.

29. What is the Kernel Trick in SVM?

Implicitly maps data to higher dimensions without computing the transformation. Enables non-linear boundaries.

30. How does Naive Bayes work?

Applies Bayes' theorem assuming features are independent given the class. Works well for text.

31. What are Decision Tree splitting criteria?

Gini: $1 - \sum p_i^2$. Entropy: $-\sum p_i \log p_i$. Lower impurity = better split.

32. How does Random Forest reduce variance?

Bagging + random feature selection at each split decorrelates trees.

33. What is Gradient Boosting?

Sequentially adds trees that predict the residual errors of previous trees.

34. What is the difference between Hard and Soft Voting?

Hard: Majority class wins. Soft: Average probabilities, pick highest.

35. What is Stacking?

Use predictions of multiple models as features for a meta-model.

Deep Learning Basics

36. What is Backpropagation?

Algorithm to compute gradients using chain rule, propagating error from output to input.

37. What is Gradient Descent?

Iteratively update weights in direction of negative gradient: $w = w - \alpha \nabla L$.

38. SGD vs Adam optimizer?

SGD: Simple, noisy. Adam: Adaptive learning rates per parameter, faster convergence.

39. What is Batch Normalization?

Normalizes layer inputs to zero mean, unit variance. Speeds up training, allows higher LR.

40. What causes Vanishing Gradients?

Sigmoid/Tanh derivatives become very small in deep nets. Fix: ReLU, ResNets, BatchNorm.

41. What is Dropout?

Randomly zeroes neurons during training with probability p. Regularization technique.

42. What are Skip Connections?

$y = F(x) + x$. Allows gradients to flow directly, enabling very deep networks.

43. What is the difference between Epoch, Batch, and Iteration?

Epoch: One pass through entire dataset. Batch: Subset of data. Iteration: One weight update.

44. Why ReLU over Sigmoid?

ReLU: No vanishing gradient (for positive), faster to compute. Sigmoid: Vanishing gradient, expensive.

45. What is Early Stopping?

Stop training when validation loss stops improving. Prevents overfitting.

Evaluation & Metrics

46. How do you handle class imbalance?

Resampling (SMOTE), class weights, different metrics (F1, PR-AUC), anomaly detection.

47. What is Stratified K-Fold?

Ensures each fold has same class distribution as original data.

48. What is Log-Loss?

Cross-entropy loss. Measures how well probabilities match true labels.

49. What is AUC-ROC?

Area under ROC curve. Measures ability to distinguish classes across all thresholds.

50. What is the Matthews Correlation Coefficient?

Balanced metric for binary classification: $\frac{TP \cdot TN - FP \cdot FN}{\sqrt{(TP+FP)(TP+FN)(TN+FP)(TN+FN)}}$.

---

Level 3: Advanced (Senior / L5+)

Architecture & Theory

51. Explain the Attention mechanism mathematically.

$Attention(Q,K,V) = Softmax(\frac{QK^T}{\sqrt{d_k}})V$. Q=Query, K=Key, V=Value. Scaled dot-product.

52. Why divide by $\sqrt{d_k}$ in attention?

Prevents dot products from growing too large, which would saturate softmax and kill gradients.

53. What is Multi-Head Attention?

Run attention h times with different projections, concatenate. Captures different relationships.

54. Encoder vs Decoder in Transformers?

Encoder: Bidirectional attention, sees all tokens. Decoder: Causal attention, sees only past.

55. BERT vs GPT?

BERT: Encoder-only, bidirectional, for understanding. GPT: Decoder-only, autoregressive, for generation.

56. What is the Reparameterization Trick?

$z = \mu + \sigma \cdot \epsilon$ where $\epsilon \sim N(0,1)$. Makes sampling differentiable in VAEs.

57. How do GANs train?

Minimax game: Generator creates fakes, Discriminator distinguishes real/fake. Both improve.

58. What is Mode Collapse in GANs?

Generator produces limited variety of outputs. Fix: Wasserstein GAN, spectral normalization.

59. What is KL-Divergence?

$D_{KL}(P||Q) = \sum P(x) \log \frac{P(x)}{Q(x)}$. Measures how one distribution differs from another.

60. What is Perplexity?

$PPL = e^{H(p,q)}$ where H is cross-entropy. Lower = model is less "confused".

Production & MLOps

61. What is Data Drift?

Input distribution changes over time. Detect with PSI, K-S test.

62. What is Concept Drift?

Relationship between features and target changes. Harder to detect.

63. What is Train-Serve Skew?

Difference between training and serving data/logic. Causes production failures.

64. How do you handle A/B testing for ML?

Split traffic, monitor business metrics, statistical significance testing, gradual rollout.

65. What is a Feature Store?

Centralized repository for storing, versioning, and serving features consistently.

66. Batch vs Real-time Inference?

Batch: Periodic, high throughput. Real-time: On-demand, low latency.

67. How do you monitor model performance in production?

Track prediction distributions, latency, error rates, business metrics, feature drift.

68. What is Model Quantization?

Reduce precision (FP32 → INT8/FP16) to speed up inference and reduce memory.

69. What is Knowledge Distillation?

Train a small "student" model to mimic a large "teacher" model's outputs.

70. What is Canary Deployment?

Release new model to small percentage of traffic, monitor, then gradually expand.

Advanced Scenarios

71. How do you debug a model that works in training but fails in production?

Check data leakage, train-serve skew, feature drift, preprocessing differences.

72. How do you handle fairness in ML?

Audit for bias, use fairness metrics (demographic parity, equalized odds), debias data/model.

73. What is Positional Bias in recommendations?

Users click top items regardless of relevance. Include position as feature in training.

74. How do you handle cold-start in recommendations?

Content-based features, popular items, ask user preferences, multi-armed bandits.

75. What is the "Exploitation vs Exploration" trade-off?

Exploit: Use current best. Explore: Try new options. Balance with epsilon-greedy, UCB, Thompson Sampling.