Choose Models for Imbalanced Data and Time-Series Forecasting
Company: Amazon
Role: Data Scientist
Category: Machine Learning
Difficulty: hard
Interview Round: Technical Screen
##### Scenario
You are asked to choose and tune models for forecasting marketplace demand and detecting fraud in a highly imbalanced dataset.
##### Question
Explain how ordinary least squares linear regression works and state its key assumptions. Compare gradient-boosted trees, random forests, and bagging; when would you prefer each? Your positive class is 0.2 % of the data. How would you handle this imbalance during model training and evaluation? Describe a full workflow for building a time-series forecasting model when seasonality and trend are present.
##### Hints
Cover data preprocessing, feature engineering, resampling/weighting, proper metrics, and cross-validation for temporal data.
Quick Answer: This question evaluates competence in selecting and tuning models for time-series forecasting with trend and seasonality and for extremely imbalanced classification, while assessing understanding of OLS assumptions, trade-offs among tree ensemble methods, imbalance handling approaches, and end-to-end forecasting workflows.
