Katkat
Back

Build House Price Model Responsibly

Capital OneData Scientist · Onsite
EasyData ScientistOnsiteMachine Learning

Last updated

This problem tests your ability to design a robust, fair, and scalable machine learning pipeline for house price prediction, emphasizing data preprocessing, model selection, and ethical oversight.

Problem Statement

You are tasked with building a house price prediction model for a major lending institution. The goal is to predict the market value of homes to assist in mortgage underwriting and risk assessment.

Data & Requirements

  • Data: You have a dataset of 500,000 property records with 80+ features, including square footage, zip code, year built, and neighborhood demographics. Some records have missing values (e.g., HVAC details) and potential outliers (e.g., luxury estates).
  • Target: The target is the numerical home sale price.
  • Performance: You must achieve high predictive accuracy while ensuring the model does not exhibit significant bias against specific neighborhoods or demographic subgroups.

Your Tasks

  1. Pipeline Design: Describe your strategy for handling missing data, categorical features (like zip codes with high cardinality), and outliers.
  2. Validation Strategy: How will you split your data to avoid leakage? Given the temporal nature of real estate data, how do you handle time-based effects?
  3. Responsible AI: How would you identify and mitigate potential biases in your model? What metrics would you use to evaluate performance across different demographic groups?

Provide a high-level architecture of your modeling approach, including the choice of metrics and the criteria you would use to determine if the model is ready for production deployment.

Hints

Hint 1

Consider how temporal features like 'year of sale' might cause data leakage if the training set contains information from the future.

Hint 2

For categorical features like zip codes, compare the trade-offs between one-hot encoding, target encoding, and entity embeddings.

Hint 3

To ensure fairness, consider evaluating 'equalized odds' or 'predictive parity' metrics across protected subgroups defined by neighborhood census data.

Follow-up questions

  • How would you explain the model's price prediction to a customer who was denied a loan based on the property valuation?
  • If the model performs well on aggregate but fails on 'starter homes' specifically, how would you diagnose and fix this?
  • What monitoring steps would you implement after the model is deployed to detect feature drift or concept drift?