← Amazon Interview Insights

Amazon·Data Scientist·Technical Phone Screen·Intermediate

Intermediate
May 2026

Summary

Amazon Data Scientist interview focused on feature engineering for a propensity model. The technical screen went pretty deep on preprocessing decisions, which I wasn't fully expecting.

Questions Asked (3)

Q1

Under what circumstances would you standardize or normalize your features before training a model?

Technical Trade-offsData Modeling
Author's notes

I knew the basics but fumbled a bit explaining when it actually matters versus when it doesn't.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying that the decision depends on the algorithm's sensitivity to feature scale and the data's distribution. Then discuss specific scenarios where standardization or normalization is beneficial, such as distance-based algorithms, gradient descent optimization, and features with different units. Finally, mention cases where it's not needed, like tree-based models, and highlight the importance of applying the same transformation to test data.

Pro tip: Emphasize that while scaling can improve model performance, it also affects interpretability; for example, coefficients in linear models become comparable but lose their original unit meaning. Always consider the trade-off and document the scaling steps for reproducibility.

1. Identify algorithm requirements

Determine if the algorithm is sensitive to feature scale, such as SVM, k-NN, neural networks, or PCA. These require scaling for optimal performance.

2. Assess feature characteristics

Check if features have different units or vastly different ranges. If so, scaling helps prevent features with larger magnitudes from dominating.

3. Consider data distribution

For normalization (e.g., min-max), consider if the data has bounded ranges or outliers. For standardization (e.g., z-score), consider if the data is approximately Gaussian.

4. Evaluate model interpretability

Decide if scaling is acceptable given potential loss of interpretability. For models where coefficients are interpreted, scaling changes the meaning.

5. Apply consistently

Ensure the same scaling parameters (mean, std or min, max) from training data are applied to validation and test data to avoid data leakage.

Key Points to Mention

  • Distance-based algorithms (e.g., k-NN, SVM) and gradient descent-based models (e.g., linear regression, neural networks) benefit from scaling.
  • Tree-based models (e.g., random forest, XGBoost) are invariant to feature scaling, so it's not required.
  • Standardization (z-score) is preferred when data follows a Gaussian distribution; normalization (min-max) is useful when data has bounded ranges or for algorithms requiring inputs between 0 and 1.
  • Scaling can affect regularization (e.g., L1/L2) by penalizing coefficients differently, so it's important for models with regularization.
  • Always fit scaler on training data only and transform test data to prevent data leakage.
  • Consider the impact on interpretability: scaling changes the units of coefficients, making them less directly interpretable.

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

Q2

You have numeric features with lots of nulls or zeros. How do you decide how to handle them?

Data ModelingTechnical Trade-offs
Author's notes

This one I actually liked.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business context and the nature of the nulls/zeros (missing vs. true zero). Then systematically evaluate imputation, transformation, or exclusion strategies based on data distribution, model requirements, and potential bias. Emphasize validation and iteration to ensure the chosen approach doesn't harm model performance or interpretability.

Pro tip: Always check if zeros are actually missing values encoded as zeros—this is common in sensor or transactional data and can drastically change your handling strategy. Document your assumptions and test multiple approaches with cross-validation to avoid silent failures.

1. Understand the Data and Business Context

Investigate why nulls/zeros exist: are they missing at random, structural, or true zeros? Consult domain experts and data documentation to determine the meaning and potential impact.

2. Assess Distribution and Missingness Patterns

Analyze the distribution of the feature and the proportion of nulls/zeros. Check if missingness correlates with other variables or the target, which could introduce bias.

3. Evaluate Handling Options

Consider deletion (if missingness is low and random), imputation (mean/median/mode, model-based, or indicator variables), or transformation (e.g., log, binning). For zeros, decide if they should be treated as valid or missing.

4. Test and Validate Impact

Experiment with different strategies using cross-validation and evaluate model performance and interpretability. Monitor for data leakage and ensure the approach aligns with business goals.

5. Document and Iterate

Document the chosen approach, rationale, and any assumptions. Continuously monitor model performance and revisit the strategy as new data arrives.

Key Points to Mention

  • Distinguish between missing values and true zeros; zeros may carry meaning (e.g., no purchase) and should not be imputed blindly.
  • Consider the mechanism of missingness (MCAR, MAR, MNAR) and its implications for bias.
  • Use indicator variables to flag imputed values, preserving information about missingness.
  • For high-cardinality or skewed features, model-based imputation (e.g., KNN, regression) may outperform simple statistics.
  • Always validate the impact of the chosen strategy on model performance and business metrics.
  • Document assumptions and maintain reproducibility for audits and collaboration.

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

Q3

If multiple features in your dataset are highly correlated with each other, how do you figure out which ones to keep?

Data ModelingTechnical Trade-offs
Author's notes

Mentioned VIF and just dropping one of a correlated pair, then brought up PCA as an option if you really want to preserve variance.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging that multicollinearity can harm model interpretability and stability, then outline a systematic process to identify and handle correlated features. Emphasize that the choice depends on the model type, business goal, and whether prediction or inference is prioritized. Conclude with a practical recommendation, such as using domain knowledge or regularization, and mention how you would validate the decision.

Pro tip: At Amazon, tie your answer to customer impact and scalability: explain how removing redundant features can reduce inference cost and latency, and mention that you'd use feature importance from a tree-based model or SHAP values to guide the decision.

1. Detect and Quantify Correlation

Compute a correlation matrix (Pearson, Spearman, or VIF) to identify pairs or groups of features with high correlation (e.g., |r| > 0.8 or VIF > 5). Visualize with a heatmap to spot patterns.

2. Assess Business and Model Impact

Determine if the correlated features are critical for business interpretation or if they are redundant. Consider the model type: linear models suffer from multicollinearity, while tree-based models are more robust but may still benefit from feature reduction.

3. Apply Selection Techniques

Use domain knowledge to keep the most interpretable or actionable feature. Alternatively, apply statistical methods like PCA, regularization (Lasso), or recursive feature elimination. For tree models, use feature importance to rank and drop less important correlated features.

4. Validate and Iterate

Evaluate the impact of feature removal on model performance using cross-validation and business metrics. Monitor for any degradation and be prepared to reintroduce features if necessary.

Key Points to Mention

  • Multicollinearity can inflate coefficient variance in linear models, making interpretation unreliable.
  • Tree-based models are less sensitive to multicollinearity but feature reduction can still improve speed and reduce overfitting.
  • Use Variance Inflation Factor (VIF) to quantify multicollinearity beyond pairwise correlation.
  • Domain knowledge is crucial: sometimes correlated features represent different business aspects and should be kept.
  • Regularization techniques like Lasso can automatically perform feature selection by shrinking coefficients.
  • Consider dimensionality reduction (PCA) when features are highly correlated and interpretability is less critical.

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