Personalizing the DoorDash Retail Store Page Experience

Luming Chen, Yuan Meng, Anthony Zhou12 Dec 2023

Problem

A retail store on a delivery marketplace carries hundreds of thousands of SKUs in thousands of categories, far more than a restaurant menu. The homepage shows themed collections stacked vertically, each a row of items that scrolls horizontally. Operators used to curate collections and fix their positions by hand, which does not scale and ignores individual needs. Shoppers also differ: some reorder the same basket weekly, others look for new or seasonal products, and their intent shifts with promotions and time of day.

Example from the article: a frequent vegan-restaurant customer should see curated vegan selections in a grocery store.

Solution in brief

DoorDash replaced manual curation with a six-stage pipeline: collection generation, collection retrieval, horizontal item ranking, item post-processing, collection ranking and collection post-processing. Models score collections and items, an item-position feature removes position bias, and maximal marginal relevance (MMR) re-ranking diversifies items and collections. Business rules adjust the final order.

Questions this case answers
  • How to rank a two-level layout (lists of lists) for each user when the catalog is too large to score in full on every page load
  • How to debias a ranker trained on clicks when the interface hides all items after a fixed position unless the user scrolls
  • How to keep a relevance-sorted list from collapsing into near-duplicates without hurting relevance too much
  • How to personalize the order of collections and the items inside them on a store homepage with hundreds of thousands of SKUs?

Tools and methods

  • Two-stage retrieval and ranking — a retrieval model picks collections per page, rankers order them.
  • Engagement models (CTR-style) — predict click, add-to-cart and conversion for items and collections.
  • Position as a training feature — separates position effect from relevance.
  • Maximal marginal relevance — re-ranks items and collections for diversity.
  • Business-rule post-processing — demotes items without photos or likely out of stock, deduplicates across collections.
Continue reading — it's free

Google or a one-time email code. No password.