← Amazon Interview Insights

Amazon·Data Scientist·Technical Phone Screen·Intermediate

Intermediate
May 2026

Summary

Technical screen for a Data Scientist role at Amazon covering a pretty broad sweep of ML fundamentals. Nothing too exotic but the breadth meant you couldn't really hide gaps anywhere.

Questions Asked (4)

Q1

How do you handle missing values in a dataset before training a model, and what's your reasoning?

Technical Trade-offsAlgorithms & Data Structures
Author's notes

I rattled off the usual stuff: mean/median imputation, indicator columns for missingness, dropping rows if the rate is high enough.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by emphasizing that there is no one-size-fits-all solution; the choice depends on the nature of the missing data (MCAR, MAR, MNAR), the proportion of missingness, and the downstream model. Then walk through a systematic decision process: diagnose, choose a strategy (deletion, imputation, or model-native handling), and validate the impact on model performance.

Pro tip: Always create a missingness indicator feature when imputing, as the pattern of missingness itself can be predictive. Also, mention that you validate imputation choices via cross-validation and compare against a baseline to ensure it actually helps.

1. Diagnose missingness

Quantify the amount and pattern of missing values per feature, and determine if missingness is random (MCAR, MAR) or informative (MNAR). Use visualizations and statistical tests to understand the mechanism.

2. Assess impact and constraints

Consider the proportion of missing data, the importance of the feature, and business constraints (e.g., real-time inference). Evaluate whether deletion is acceptable or if imputation is necessary.

3. Choose a handling strategy

Select from deletion (listwise/pairwise), simple imputation (mean/median/mode), model-based imputation (KNN, MICE), or model-native handling (XGBoost, LightGBM). Justify based on data type and model assumptions.

4. Implement and validate

Apply the chosen method within a cross-validation pipeline to avoid data leakage. Compare model performance with and without imputation, and consider adding missingness indicators.

5. Monitor and iterate

After deployment, monitor for changes in missingness patterns and retrain or adjust the imputation strategy as needed. Document the rationale for future reference.

Key Points to Mention

  • Types of missingness: MCAR, MAR, MNAR and their implications
  • Trade-offs between deletion and imputation (bias vs. variance, data loss)
  • Simple vs. advanced imputation methods (mean/median vs. MICE, KNN)
  • Model-native handling (e.g., XGBoost's sparsity-aware split finding)
  • Importance of avoiding data leakage by fitting imputers only on training folds
  • Adding missingness indicator features to capture informative missingness

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

Q2

Given a business problem, walk through how you'd pick an ML algorithm and justify that choice.

Technical Trade-offsProduct Analytics & Metrics
Author's notes

This is where I rambled a bit.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business problem and translating it into an ML formulation (e.g., classification, regression, ranking). Then walk through a structured evaluation of candidate algorithms based on data characteristics, constraints, and business metrics, and justify your final choice with trade-offs.

Pro tip: Always tie your algorithm choice back to the business impact and Amazon's leadership principles, such as Customer Obsession and Dive Deep. Mention that you would start with a simple baseline (e.g., logistic regression) before moving to complex models, and validate with offline metrics that correlate with online business metrics.

1. Understand the Business Problem

Clarify the problem, success metrics, and constraints (e.g., latency, interpretability, budget). Translate it into an ML task (e.g., binary classification, ranking).

2. Assess Data and Requirements

Evaluate data size, quality, features, and label availability. Identify key requirements like real-time inference, explainability, or scalability.

3. Shortlist Candidate Algorithms

Based on the problem type and constraints, list 2-3 suitable algorithms (e.g., logistic regression, gradient boosting, neural networks). Consider baselines and state-of-the-art.

4. Evaluate Trade-offs

Compare candidates on performance (offline metrics), interpretability, training/inference time, scalability, and maintenance. Consider business metrics like ROI or customer satisfaction.

5. Justify and Iterate

Select the best algorithm with clear reasoning, and propose an experimentation plan (A/B test) to validate. Mention that you would iterate based on results.

Key Points to Mention

  • Problem formulation: mapping business problem to ML task (classification, regression, ranking, etc.)
  • Data characteristics: size, dimensionality, sparsity, label quality, and feature types
  • Algorithm trade-offs: bias-variance, interpretability vs. accuracy, training/inference latency, scalability
  • Business metrics alignment: offline metrics (AUC, RMSE) vs. online metrics (CTR, conversion, revenue)
  • Baseline models: starting simple (e.g., logistic regression) before complex models
  • Experimentation: A/B testing and iterative improvement based on business impact

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

Q3

Can you explain how Random Forests work in plain language and how they differ from linear regression?

Algorithms & Data StructuresTechnical Trade-offs
Author's notes

Honestly the contrast part tripped me up more than explaining Random Forests.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start with a plain-language analogy for Random Forests, then contrast with linear regression on assumptions, interpretability, and performance. Use a concrete example to illustrate when each shines, and tie it back to real-world data science trade-offs.

Pro tip: Emphasize that Random Forests capture non-linear relationships and interactions without explicit feature engineering, but linear regression offers interpretable coefficients and faster inference—choose based on the problem's need for explainability versus predictive power.

1. Define Random Forests Simply

Explain that a Random Forest is an ensemble of many decision trees, each trained on a random subset of data and features, and their predictions are averaged (regression) or voted (classification).

2. Define Linear Regression Simply

Describe linear regression as fitting a straight line (or hyperplane) that models the relationship between inputs and a continuous output by minimizing squared errors.

3. Contrast Key Differences

Highlight differences in assumptions (linearity vs. non-linearity), interpretability (coefficients vs. black-box), and handling of interactions and outliers.

4. Discuss Trade-offs and Use Cases

Explain when to use each: linear regression for simple, interpretable models with linear relationships; Random Forests for complex, non-linear data where accuracy matters more than explainability.

5. Tie to Amazon Context

Mention how these trade-offs apply to real-world scenarios like demand forecasting or customer segmentation, where both interpretability and predictive power are valued.

Key Points to Mention

  • Ensemble learning and bagging in Random Forests
  • Assumption of linearity in linear regression
  • Interpretability: coefficients vs. feature importance
  • Handling of non-linear relationships and interactions
  • Overfitting: Random Forests are less prone but can overfit with deep trees; linear regression can overfit with many features
  • Computational cost: Random Forests are slower to train and predict, linear regression is fast

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

Q4

What's the difference between overfitting and underfitting, and how do you detect and address each?

Algorithms & Data StructuresA/B Testing & Experimentation
Author's notes

Felt most comfortable here.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Define overfitting and underfitting clearly, then explain how to detect each using learning curves and validation metrics, and finally describe practical techniques to address them. Emphasize the bias-variance tradeoff and how you would apply these concepts in a real-world Amazon data science context.

Pro tip: Mention that in practice, you often start by checking if the model is overfitting or underfitting using learning curves, and then apply regularization or more data accordingly. Also, highlight that underfitting can sometimes be addressed by increasing model complexity or feature engineering, while overfitting requires simplifying the model or adding more data.

1. Define the concepts

Clearly explain overfitting (model learns noise, high variance) and underfitting (model too simple, high bias) with examples.

2. Detection methods

Describe how to detect each using learning curves, validation metrics, and comparing training vs. validation performance.

3. Addressing overfitting

List techniques like regularization (L1/L2), dropout, early stopping, cross-validation, and gathering more data.

4. Addressing underfitting

List techniques like increasing model complexity, adding features, reducing regularization, and training longer.

5. Trade-off and practical application

Discuss the bias-variance tradeoff and how to balance both in a real project, possibly with an example from A/B testing or experimentation.

Key Points to Mention

  • Bias-variance tradeoff
  • Learning curves and validation curves
  • Regularization techniques (L1, L2, dropout)
  • Cross-validation and early stopping
  • Feature engineering and model complexity
  • Impact on A/B testing and experimentation (e.g., overfitting to historical data can lead to false positives in experiments)

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