AWS MLE

Question 1

What is regularization in machine learning and why is it important?

Answer 1

Regularization is the addition of a penalty term to a model’s loss function to discourage overfitting by limiting model complexity. Common methods include L1 (lasso) for feature selection, L2 (ridge) for weight shrinkage, and dropout in neural networks.

Question 2

When would you prefer L1 over L2 regularization?

Answer 2

Use L1 when feature sparsity and interpretability are desired, since it drives some coefficients exactly to zero. L2 is better for stability and when all features contribute small effects.

Question 3

What is model drift and what types are there?

Answer 3

Model drift is the degradation of model performance over time as data distribution changes. Two main types:

Data drift (covariate shift): Input feature distribution changes.

Concept drift: Relationship between input and target changes.

Question 4

How can you detect and mitigate model drift?

Answer 4

Monitor prediction quality and statistical distributions of inputs/outputs, set alerts for significant changes, retrain on updated data, use online learning or scheduled retraining pipelines.

Question 5

What are common strategies for hyperparameter tuning?

Answer 5

Grid search, random search, Bayesian optimization, Hyperband, population-based training. Random search often outperforms grid for high-dimensional spaces.

Question 6

How does Bayesian optimization improve tuning efficiency?

Answer 6

It builds a surrogate model (like Gaussian Process) of the objective function, then uses acquisition functions (e.g., expected improvement) to balance exploration vs. exploitation, reducing number of evaluations needed.

Question 7

What are the key components of a scalable ML system design?

Answer 7

1. Data ingestion & preprocessing
2. Feature store
3. Training pipeline
4. Model registry
5. Deployment endpoints
6. Monitoring (data quality, drift, performance)
7. Retraining/CI-CD integration.

Question 8

What role does a feature store play in ML systems?

Answer 8

It centralizes, versions, and serves consistent features for training and inference, reducing data leakage and ensuring reproducibility.

Question 9

Why is model registry important?

Answer 9

It stores model versions, metadata, performance metrics, and lineage, enabling reproducible deployments, auditing, and rollback if needed.

Question 10

How would you design monitoring for a real-time fraud detection system?

Answer 10

Collect prediction + input logs, monitor latency and throughput, implement drift detection on inputs, track precision/recall using delayed ground truth, and trigger retraining when thresholds are exceeded.

Question 11

What are the main stages of a SageMaker pipeline?

Answer 11

1. Data preprocessing (Processing Jobs)
2. Training (Training Jobs)
3. Hyperparameter tuning (HPO Jobs)
4. Model evaluation
5. Model registration (Model Registry)
6. Deployment (Endpoints/Batches)
7. Monitoring (Model Monitor).

Question 12

How does SageMaker simplify ML workflow automation?

Answer 12

It provides managed services for data processing, distributed training, automatic scaling, experiment tracking, built-in algorithms, and integrates with CI/CD via SageMaker Pipelines and Step Functions.

Question 13

Why is data versioning critical in ML?

Answer 13

It ensures reproducibility, traceability, and auditability of experiments—allowing you to link model artifacts to the exact dataset and features used.

Question 14

What are common tools and approaches for data versioning?

Answer 14

DVC, LakeFS, MLflow with artifact tracking, Delta Lake for time travel, or AWS solutions like S3 versioning + SageMaker Lineage Tracking.

Question 15

What is the bias–variance tradeoff

Answer 15

It describes the balance between underfitting and overfitting. High bias models are too simple and underfit, while high variance models are too complex and overfit. The goal is to minimize total error by balancing both.

Question 16

Give an example of a high-bias algorithm and a high-variance algorithm.

Answer 16

High-bias: Linear regression on a non-linear dataset.
High-variance: Deep decision trees without pruning or regularization.

Question 17

How can you reduce variance without increasing bias too much?

Answer 17

Use regularization (L1/L2, dropout), ensemble methods (bagging, random forests), cross-validation, or more training data.

Question 18

How can you reduce bias in a model?

Answer 18

se a more complex model (e.g., deeper neural nets, more features), reduce regularization strength, or engineer more informative features.

Question 19

What is precision and what is recall?

Answer 19

Precision = TP / (TP + FP), the fraction of positive predictions that are correct.

Recall = TP / (TP + FN), the fraction of actual positives correctly identified.

Question 20

What is the precision–recall tradeoff?

Answer 20

Increasing precision usually decreases recall, and vice versa. Adjusting the decision threshold shifts the balance. The optimal point depends on the problem’s cost of false positives vs. false negatives.

Question 21

When would you prioritize recall over precision?

Answer 21

When missing a positive case is more costly than flagging false positives. Example: detecting cancer or fraud.

Question 22

When would you prioritize precision over recall?

Answer 22

When false positives are very costly or harmful. Example: recommending a financial loan approval or automated hiring filters.

Question 23

How does the PR curve differ from the ROC curve?

Answer 23

PR curve plots precision vs. recall; it’s more informative for imbalanced datasets. ROC curve plots TPR vs. FPR, but can be misleading when negatives vastly outnumber positives.

Question 24

What is the F1 score and why is it useful?

Answer 24

The harmonic mean of precision and recall (2 × P × R / (P + R)), useful when you want a balance and the class distribution is skewed.

Question 25

Why do we use the harmonic mean in F1 score?

Answer 25

We use the harmonic mean (instead of the arithmetic mean) because it penalizes extreme imbalance between precision and recall. Arithmetic mean could give a deceptively high score if one metric is very large and the other is very small. Example: Precision = 1.0, Recall = 0.0 → Arithmetic mean = 0.5 (misleadingly high). Harmonic mean emphasizes the lower value — if either precision or recall is near zero, the F1 score will also be near zero. Same example: Precision = 1.0, Recall = 0.0 → F1 = 0.0 (correctly reflecting poor performance). So the harmonic mean ensures that a high F1 requires both precision and recall to be reasonably high, making it a better measure when balancing both.

Question 26

How would you interpret coefficients of logistic regression for categorical and boolean variables?

Answer 26

Focus on interpreting odds ratios, the meaning of coefficients relative to a baseline for categorical variables, and how boolean inputs affect log-odds. Be clear about dummy variable encoding and interaction effects. Use real-world analogies if possible, like conversion rates or click-through probabilities.

Question 27

Let’s say we want to build a model to predict booking prices. Between linear regression and random forest regression, which model would perform better and why?

Answer 27

Start by comparing assumptions—linear regression needs linear relationships, random forest doesn’t. Factor in interpretability vs. performance, training data size, and handling of outliers or skewed distributions. At Amazon, model choice often balances explainability (for stakeholders) with performance (for automated decisions).

Question 28

Say you are given a dataset of perfectly linearly separable data. What would happen when you run logistic regression?

Answer 28

Tests edge-case awareness in optimization. Explain how logistic regression fails to converge in this case, as it endlessly increases weights to find a “better” separation. Discuss how regularization reintroduces a trade-off, allowing convergence. Useful for scenarios involving imbalanced or clean datasets—common in anomaly detection or QA tagging.

Question 29

What is a K-means algorithm?

Answer 29

K-means is an unsupervised clustering algorithm that partitions data into k clusters. It iteratively assigns points to the nearest cluster centroid, then updates centroids as the mean of assigned points, until convergence. It minimizes within-cluster variance.

Question 30

What is the difference between SVM and logistic regression?

Answer 30

Logistic regression is probabilistic and linear, optimizing log-loss to estimate class probabilities. SVM is margin-based, focusing on finding the maximum margin hyperplane. SVM can use kernels to model non-linear boundaries, while logistic regression remains linear unless features are engineered.

Question 31

Describe normalization and Bayes’ rule.

Answer 31

Normalization rescales data to a common scale (e.g., min-max or z-score) to improve training stability. Bayes’ rule relates conditional probabilities: P(A|B) = [P(B|A) * P(A)] / P(B), useful in Bayesian models and Naïve Bayes classifiers.

Question 32

Describe linear regression versus logistic regression.

Answer 32

Linear regression predicts continuous outputs using least squares. Logistic regression predicts probabilities of categorical outcomes using a sigmoid function and cross-entropy loss. Linear assumes Gaussian residuals; logistic assumes Bernoulli outcomes.

Question 33

What kind of different loss functions do you know?

Answer 33

Regression: MSE, MAE, Huber loss. Classification: Cross-entropy, hinge loss, focal loss. Ranking: Pairwise loss, NDCG loss. Computer vision: IoU loss, Dice loss, perceptual loss. Regularization: L1, L2.

Question 34

How do you measure the performance of computer vision models?

Answer 34

Use accuracy, precision/recall, F1, IoU (Intersection over Union), mAP (mean Average Precision), top-k accuracy, or confusion matrices depending on task (classification, detection, segmentation).

Question 35

How do you deal with a troublesome dataset?

Answer 35

Diagnose issues (missing values, noise, outliers, imbalance), apply preprocessing (imputation, normalization, augmentation), perform feature engineering, and consult domain experts. For persistent issues, re-collect or refine labeling.

Question 36

How do you deal with misrepresentative training data (imbalanced dataset, overfitting, explain how L1/L2 regularization work at an optimization level)?

Answer 36

Imbalanced data: resampling, class weights, anomaly detection framing. Overfitting: dropout, weight decay (L2), early stopping, data augmentation. L1 regularization (adds |w| term): promotes sparsity by pushing weights to zero. L2 regularization (adds w² term): shrinks weights smoothly, distributing importance. Both modify the optimization objective by penalizing large weights.