← Amazon Interview Insights

Amazon·Data Scientist·Technical Phone Screen·Intermediate

Intermediate
Jun 2026

Summary

Rapid whiteboard session for a Data Scientist role at Amazon, covering a broad sweep of supervised learning fundamentals in quick succession. The whole thing felt like a pop quiz more than a conversation, which threw me off a bit.

Questions Asked (7)

Q1

Explain the bias-variance trade-off and how you would diagnose whether a model is suffering from one or the other.

Technical Trade-offsProduct Analytics & Metrics
Author's notes

I knew this cold but still fumbled the diagnosis part.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clearly defining bias and variance and the trade-off between them, then explain how to diagnose each using learning curves and validation metrics. Emphasize practical steps to address the issue once diagnosed, and tie it back to business impact.

Pro tip: Mention that in practice, you often need to balance bias and variance with business constraints, and that sometimes a slightly biased model is preferable if it's more interpretable or faster. Also, highlight that Amazon values customer obsession, so relate the trade-off to delivering value to customers.

1. Define Bias and Variance

Explain bias as error from erroneous assumptions (underfitting) and variance as sensitivity to fluctuations in the training set (overfitting).

2. Explain the Trade-off

Describe how increasing model complexity reduces bias but increases variance, and vice versa, leading to a U-shaped test error curve.

3. Diagnose Bias

Use learning curves: if both training and validation errors are high and converge, the model is biased. High training error indicates underfitting.

4. Diagnose Variance

If training error is low but validation error is high, and there's a large gap, the model is overfitting (high variance).

5. Address the Issue

For bias: increase model complexity, add features, reduce regularization. For variance: get more data, reduce features, increase regularization, use ensemble methods.

Key Points to Mention

  • Bias-variance decomposition of expected test error
  • Learning curves for diagnosis
  • Underfitting vs overfitting symptoms
  • Regularization techniques (L1/L2, dropout)
  • Ensemble methods (bagging, boosting) to reduce variance
  • Cross-validation for reliable error estimates

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q2

What are the key assumptions behind ordinary least-squares linear regression?

Technical Trade-offs
Author's notes

Listed the usual suspects: linearity, independence, homoscedasticity, normality of residuals, no multicollinearity.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining OLS as a method that minimizes the sum of squared residuals, then systematically list the key assumptions (linearity, independence, homoscedasticity, normality, no multicollinearity, no endogeneity). For each assumption, briefly explain its importance and the consequences of violation, and mention how to check it. Conclude by noting that while some assumptions are critical for unbiasedness and inference, others affect efficiency or standard errors.

Pro tip: Emphasize that in practice, the Gauss-Markov assumptions (linearity, exogeneity, homoscedasticity, no autocorrelation) are crucial for OLS to be BLUE, while normality is only needed for small-sample inference. This shows you understand the hierarchy of assumptions and can prioritize them in real-world modeling.

1. Define OLS and its goal

Briefly state that OLS estimates coefficients by minimizing the sum of squared residuals, aiming to find the best linear unbiased estimator (BLUE) under certain conditions.

2. List the core assumptions

Enumerate the key assumptions: linearity in parameters, exogeneity (zero conditional mean of errors), homoscedasticity (constant error variance), no autocorrelation (independent errors), normality of errors (for inference), and no perfect multicollinearity.

3. Explain consequences of violations

For each assumption, describe what happens if violated: e.g., omitted variable bias from exogeneity violation, inefficient estimates from heteroscedasticity, invalid standard errors from autocorrelation, etc.

4. Discuss diagnostics and remedies

Mention common checks (residual plots, Breusch-Pagan test, Durbin-Watson, VIF) and potential fixes (transformations, robust standard errors, adding variables, regularization).

5. Summarize and prioritize

Conclude by highlighting which assumptions are most critical for unbiasedness (exogeneity, linearity) versus those for valid inference (normality, homoscedasticity) and note that OLS can still be useful even if some assumptions are mildly violated.

Key Points to Mention

  • Linearity: The relationship between predictors and outcome is linear in parameters.
  • Exogeneity: The error term has zero conditional mean (no omitted variable bias, no simultaneity).
  • Homoscedasticity: Constant variance of errors across all levels of predictors.
  • No autocorrelation: Errors are uncorrelated across observations (important for time series).
  • Normality of errors: Errors are normally distributed (needed for small-sample hypothesis tests).
  • No perfect multicollinearity: Predictors are not perfectly correlated (otherwise coefficients are not uniquely estimable).

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q3

How do you detect overfitting and what steps do you take to address it?

Technical Trade-offsRoot Cause Analysis
Author's notes

Went with cross-validation, regularization, and early stopping.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining overfitting and explaining why it's a critical issue in machine learning. Then, describe a systematic process for detecting and addressing it, using concrete examples from your experience. Emphasize the importance of validation techniques and iterative model improvement.

Pro tip: Quantify the impact of overfitting on business metrics and show how your approach mitigates it, aligning with Amazon's customer obsession and bias for action.

1. Define Overfitting

Explain what overfitting is: a model that performs well on training data but poorly on unseen data, capturing noise rather than signal.

2. Detection Methods

Describe techniques to detect overfitting, such as monitoring training vs. validation error, using learning curves, and employing cross-validation.

3. Diagnose Causes

Identify potential causes: model complexity, insufficient data, noisy features, or data leakage. Use tools like learning curves to diagnose.

4. Mitigation Strategies

List steps to address overfitting: simplify model, gather more data, feature selection, regularization, early stopping, dropout, data augmentation, and ensemble methods.

5. Validate and Iterate

Emphasize the need to re-evaluate after mitigation, using a hold-out test set or cross-validation, and iterate until performance generalizes well.

Key Points to Mention

  • Training vs. validation error gap
  • Cross-validation techniques (k-fold, stratified)
  • Regularization methods (L1, L2, dropout)
  • Early stopping and model complexity control
  • Data augmentation and feature engineering
  • Ensemble methods (bagging, boosting)

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q4

Compare boosting and bagging as ensemble methods, and describe when you would prefer one over the other.

Technical Trade-offsSystem Design
Author's notes

This is where I actually felt decent.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining boosting and bagging, highlighting their core differences in bias-variance trade-off and training approach. Then, discuss when to prefer each based on data characteristics, computational resources, and business goals, using examples to illustrate. Conclude with a practical recommendation for a scenario like Amazon's.

Pro tip: Emphasize that boosting often requires careful hyperparameter tuning and is sensitive to noisy data, while bagging is more robust and parallelizable. Mention that in practice, XGBoost (boosting) often wins competitions but bagging (e.g., Random Forest) is preferred for quick, reliable baselines.

1. Define the methods

Briefly explain bagging (bootstrap aggregating) and boosting (sequential ensemble) with examples like Random Forest and AdaBoost/XGBoost.

2. Compare key characteristics

Contrast them in terms of bias-variance reduction, training parallelism, sensitivity to noise, and computational cost.

3. Discuss when to prefer each

Outline scenarios: bagging for high-variance models and noisy data; boosting for high accuracy and low bias, with sufficient data and tuning.

4. Relate to business context

Tie the choice to practical considerations like interpretability, scalability, and deployment constraints, especially in a company like Amazon.

5. Summarize with a recommendation

Provide a concise recommendation based on the trade-offs, showing awareness of both theoretical and practical aspects.

Key Points to Mention

  • Bias-variance trade-off: Bagging reduces variance, boosting reduces bias.
  • Training parallelism: Bagging can be parallelized; boosting is sequential.
  • Sensitivity to noisy data and outliers: Boosting more sensitive; bagging more robust.
  • Computational cost and tuning: Boosting often requires more tuning and resources.
  • Examples: Random Forest (bagging) vs. XGBoost/AdaBoost (boosting).
  • Use cases: Bagging for high-variance models and noisy data; boosting for high accuracy and structured data.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q5

Which evaluation metrics would you select for a binary classifier with heavily imbalanced classes, and why?

Product Analytics & MetricsTechnical Trade-offs
Author's notes

Precision-recall and AUC-PR over ROC-AUC because ROC can look deceptively good when negatives dominate.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging that accuracy is misleading for imbalanced data, then recommend metrics that focus on the minority class such as precision, recall, F1, and AUC-PR. Explain the trade-offs between these metrics and how they align with business objectives, and mention that the choice depends on the specific costs of false positives and false negatives.

Pro tip: At Amazon, tie your metric choice to customer impact and business metrics—for example, in fraud detection, recall is critical to avoid customer friction, but precision matters to reduce false alarms. Show you understand the operational trade-offs.

1. Explain why accuracy fails

State that with imbalanced classes, a naive model predicting the majority class can achieve high accuracy but poor minority class performance, making accuracy misleading.

2. Introduce precision, recall, and F1

Define precision (positive predictive value) and recall (sensitivity) and explain that F1 balances them. Discuss how the choice depends on whether false positives or false negatives are more costly.

3. Discuss threshold-independent metrics

Mention AUC-ROC and AUC-PR. Explain that AUC-PR is more informative for imbalanced data because it focuses on the minority class, while AUC-ROC can be overly optimistic.

4. Consider business context and costs

Align metric selection with business objectives. For example, in fraud detection, high recall may be prioritized to catch fraud, but precision is also important to avoid blocking legitimate transactions.

5. Recommend a combination and monitoring

Suggest using a combination of metrics (e.g., AUC-PR, recall at a fixed precision, F1) and monitoring them over time. Also consider using techniques like resampling or class weights, and evaluate their impact on metrics.

Key Points to Mention

  • Accuracy paradox: high accuracy can be achieved by predicting the majority class only.
  • Precision and recall trade-off: depends on the relative cost of false positives vs. false negatives.
  • F1 score: harmonic mean of precision and recall, useful when you need a balance.
  • AUC-PR (Precision-Recall AUC): more informative than AUC-ROC for imbalanced datasets.
  • Business alignment: metrics should reflect the actual costs and goals of the application.
  • Threshold tuning: choose a threshold that optimizes the desired metric for the business.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q6

How would you handle missing data in your training set?

Data ModelingTechnical Trade-offs
Author's notes

Covered imputation strategies and flagging missingness as a feature.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by emphasizing that the approach depends on the nature and amount of missing data, and that you would first diagnose the missingness mechanism. Then outline a systematic process: analyze patterns, choose appropriate handling methods (deletion, imputation, or model-based), and validate the impact on model performance. Highlight the importance of aligning with business context and avoiding data leakage.

Pro tip: At Amazon, always tie your technical choices to customer impact and scalability—e.g., how missing data handling affects model reliability at scale. Also, mention that you would document and monitor missing data patterns in production to detect drift.

1. Diagnose Missingness

Determine the mechanism (MCAR, MAR, MNAR) and quantify the extent of missing data per feature. Use visualizations and statistical tests to understand patterns.

2. Assess Impact and Business Context

Evaluate how missing data affects the target variable and business objectives. Consider if missingness itself is informative (e.g., a missing value could indicate a specific customer behavior).

3. Choose Handling Strategy

Select from deletion, simple imputation (mean/median/mode), advanced imputation (KNN, MICE, regression), or model-based methods (XGBoost handles missing natively). Consider adding missing indicator features.

4. Validate and Iterate

Use cross-validation to compare model performance with different imputation methods. Ensure no data leakage by fitting imputers only on training folds.

5. Monitor in Production

Set up monitoring for missing data rates and patterns to detect drift. Have a fallback strategy if missingness increases unexpectedly.

Key Points to Mention

  • Missing data mechanisms: MCAR, MAR, MNAR and their implications
  • Trade-offs between deletion (loss of information) and imputation (potential bias)
  • Advanced imputation techniques like MICE, KNN, or using algorithms that handle missing values natively (e.g., XGBoost)
  • Creating missing indicator features to capture informative missingness
  • Avoiding data leakage by fitting imputation only on training data
  • Evaluating the impact on model performance and business metrics

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q7

Why does the normal distribution appear so frequently in statistical modeling?

Technical Trade-offs
Author's notes

Central limit theorem, maximum entropy under mean and variance constraints.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging the mathematical foundation (Central Limit Theorem) and then explain why it's practically useful in modeling (e.g., noise, aggregation, and conjugate priors). Finally, connect it to real-world data science scenarios, especially at Amazon, where normal assumptions often simplify complex problems.

Pro tip: Mention that while the normal distribution is ubiquitous, it's crucial to check assumptions (e.g., normality tests, Q-Q plots) and consider alternatives like heavy-tailed distributions when data violates normality, showing you understand both theory and practice.

1. Mathematical Foundation

Explain the Central Limit Theorem: sums/averages of independent random variables tend toward normality, making it a natural limit for many processes.

2. Practical Convenience

Highlight mathematical tractability: closed-form solutions, easy parameter estimation, and conjugate priors in Bayesian inference.

3. Real-World Occurrence

Discuss how measurement errors, natural phenomena, and aggregated metrics often approximate normality due to many small, independent effects.

4. Role in Statistical Modeling

Mention its use in linear regression (Gaussian errors), hypothesis testing (t-tests, ANOVA), and as a default assumption in many models.

5. Caveats and Alternatives

Acknowledge that real data can be skewed or heavy-tailed, and that blindly assuming normality can lead to poor models; suggest checking and transforming data.

Key Points to Mention

  • Central Limit Theorem
  • Maximum entropy principle (normal distribution is the most uncertain for given mean and variance)
  • Conjugate priors in Bayesian statistics (normal-normal, normal-inverse-gamma)
  • Gaussian assumptions in linear models and error terms
  • Common in natural and social sciences due to aggregation of small effects
  • Importance of validating normality assumptions and considering robust alternatives

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.