Katkat
Back

What features and feature selection would you use?

MetaData Scientist · Technical Screen
MediumData ScientistTechnical ScreenMachine Learning

Last updated

This question evaluates feature engineering, feature selection, and production-aware machine learning system design skills for ranking shop ads, within the Machine Learning domain for a Data Scientist role.

Context

You are building an ML system to rank/promote shop ads in an e-commerce feed/search page. At serving time, the system may score candidate shop ads for a given user and context.

Assume you have access to:

  • User events (impressions, clicks, purchases, shop follows)
  • Shop metadata (category, price bands, inventory signals)
  • Query/context (search query, time, device)
  • Ad/auction signals (bid, budget pacing)

Questions

  1. If you were to build the shop-ads ranking model, what feature families would you use? (Give examples.)
  2. You have “a ton” of candidate features. How would you identify which ones are useful?
  • Include at least one offline approach and one online/production-safe approach.
  1. If you were not allowed to use a model-based importance method (e.g., no SHAP/GBDT gain/permutation importance), how would you still find the key useful features?
  2. Call out common pitfalls: leakage, feedback loops, cold start, and feature drift.