How would you design a recommendation system for a new e-commerce platform?

Example Answer

I would start with the business objective because recommendations can optimize very different things: conversion, revenue, discovery, repeat purchase, or long-term customer value. That choice affects everything from label design to ranking logic.

For a new platform, cold start is the first real problem, so I would not begin with a purely collaborative approach. I would use content and catalog features for initial candidate generation, layer in popularity and business rules where needed, and then add behavioral signals as interaction data accumulates. Over time I would separate the system into candidate generation and ranking so it can scale, personalize, and support different surfaces such as home page, PDP, cart, and email.

I would also evaluate beyond CTR. Recommendations should be judged on quality, diversity, coverage, and downstream business value, otherwise the system can become a narrow click optimizer that hurts the broader shopping experience.

From data to a deployed baseline

Collect permitted catalog attributes, availability and price, plus impression logs that identify what was actually shown, where and when. Join clicks and purchases to those exposures with a defined observation window; an item never shown is not automatically a negative preference. Start with eligible popular/category items and content similarity. As labels accumulate, compare a simple logistic or boosted-tree ranker on user, item and context features with that baseline using later time periods and cold-start slices.

Serve candidate retrieval and ranking behind a deadline, filter unavailable or ineligible products, and fall back to the baseline when personalized features are missing. Log the model/version and exposure policy. Test changes with stable assignment and downstream purchase/return and latency guardrails; offline clicks inherit the previous policy's exposure bias. Add more complex retrieval only when observed quality and serving cost justify it.

Common Poor Answer

A weak answer says "use collaborative filtering" without dealing with cold start, surface-specific objectives, or how the system will be measured beyond clicks.