Optimize Feature Selection and Handling in Machine Learning Models
Quick Overview
Evaluates practical feature engineering for customer propensity models. Strong answers explain when to scale variables, how to handle missing and zero-heavy numeric predictors, and how to deal with correlated features while balancing model family, leakage, interpretability, and production reliability.
Optimize Feature Selection and Handling in Machine Learning Models
Company: Amazon
Role: Data Scientist
Category: Machine Learning
Difficulty: medium
Interview Round: Technical Screen
##### Scenario
Feature engineering for a customer-propensity machine-learning model.
##### Question
When do we need to standardize or normalize variables? How would you handle numeric predictors that contain many null or zero values? If several features are highly correlated, how would you decide which one(s) to keep?
##### Hints
Discuss scaling impact, imputation vs flagging, and multicollinearity remedies.
Quick Answer: Evaluates practical feature engineering for customer propensity models. Strong answers explain when to scale variables, how to handle missing and zero-heavy numeric predictors, and how to deal with correlated features while balancing model family, leakage, interpretability, and production reliability.
Optimize Feature Selection and Handling in Machine Learning Models
You are building a customer propensity model to predict whether a user will purchase, subscribe, or take another target action. The data includes transaction, app/web activity, demographic, and marketing features.
Constraints & Assumptions
Treat this as a practical feature-engineering question for production modeling.
Discuss how choices depend on the algorithm, not just the data type.
Include leakage, missingness, zero inflation, correlation, interpretability, and monitoring considerations.
Assume the target and prediction time are already defined.
Clarifying Questions to Ask Guidance
Which model families are being considered: linear, tree-based, distance-based, neural, or ensemble?
Are missing values meaningful signals or data quality artifacts?
Do zeros represent true zero behavior, missing data encoded as zero, or a separate state?
Is the priority prediction accuracy, interpretability, or operational simplicity?
Part 1 - Scaling Variables
When do we need to standardize or normalize variables?
What This Part Should Cover Guidance
Why scaling matters for distance-based models, gradient-based models, regularized linear models, PCA, and neural networks.
Why tree-based models usually do not require scaling for predictive performance.
Standardization versus min-max normalization and when each is useful.
Part 2 - Missing and Zero-Heavy Numeric Features
How would you handle numeric predictors that contain many null or zero values?
What This Part Should Cover Guidance
Data-quality checks to distinguish missing, not applicable, and true zero.
Imputation choices, missingness indicators, zero flags, bucketing, transformations, and model families that handle missing values.
Zero-inflated behavior and whether zero itself is predictive.
Train-serving consistency and leakage prevention.
Part 3 - Highly Correlated Features
If several features are highly correlated, how would you decide which ones to keep?
What This Part Should Cover Guidance
Correlation analysis, VIF, feature importance, domain meaning, redundancy, stability, and cost to compute.
Regularization, feature selection, dimensionality reduction, grouping, or keeping correlated features for tree ensembles when helpful.
Impact on interpretability versus predictive accuracy.
What a Strong Answer Covers Guidance
A strong answer makes feature-handling decisions conditional on model family and business use case, checks data meaning before transforming, and balances predictive performance, interpretability, leakage risk, and production reliability.
Follow-up Questions Guidance
How would you detect a feature that leaks the target?
What if a zero-heavy feature is the strongest predictor?
