Katkat
Back

Design a house-price prediction model

Two SigmaData Scientist · Technical Screen
EasyData ScientistTechnical ScreenMachine Learning

Last updated

This problem evaluates your ability to design a robust end-to-end machine learning pipeline for house price prediction, focusing on data preprocessing, feature engineering, and validation strategy.

Scenario

You are tasked with building a regression model to predict the sale price of residential properties. You have access to a dataset containing 100,000 house sales over the past five years. Key features include square footage, number of bedrooms/bathrooms, year built, neighborhood (categorical), and the historical sale price.

Task

Design a machine learning pipeline to address the following requirements:

  1. Data Preprocessing: How would you handle missing values (e.g., missing basement area) and high-cardinality features like 'Neighborhood'?
  2. Feature Engineering: Propose three specific features you would derive to improve model performance, explaining the intuition behind each.
  3. Validation Strategy: Given that housing markets change over time, how would you split your data for training and testing to ensure the model generalizes to future market conditions?
  4. Metrics: Which evaluation metric would you choose to measure performance, and why might you prefer it over Mean Squared Error (MSE)?

Example Case

Suppose a new data point arrives: A house in an affluent, up-and-coming neighborhood with 2,500 sqft, built in 1990, with a missing value for the 'lot size' column. Briefly describe how your pipeline handles the missing lot size and the categorical neighborhood encoding before it reaches the model.

Constraints

  • Assume the model must be interpretable for real estate agents.
  • Be prepared to discuss how you would identify and treat outliers in the price distribution.

Hints

Hint 1

Consider how temporal leakage might occur if you use a simple random train-test split on time-series data.

Hint 2

For high-cardinality categorical features, think about methods beyond one-hot encoding, such as target encoding or embeddings.

Hint 3

Think about the difference between absolute error metrics like MAE and squared error metrics like MSE, especially regarding the impact of outliers on price prediction.

Follow-up questions

  • How would you detect and account for 'data drift' where market trends shift significantly after your model is deployed?
  • If the model shows high variance (overfitting), what specific regularization or architectural changes would you implement?
  • How would you explain the impact of 'square footage' on the prediction to a non-technical stakeholder using model interpretability tools?