Problem
A fashion retailer has about 100,000 items in the catalogue at any time, and hundreds of new items arrive every week. Customers meet products on category pages, search results, product pages and a home feed. Category and search pages can hold thousands of products, and the "You Might Also Like" carousel has only 16 slots. Each surface needs its own personalised ranking, and it must be fast during sale peaks. New items have no interactions to learn from.
Example from the article: ASOS ranks products by clicks and purchases on four surfaces: category page, "You Might Also Like" carousel, "For You" tab on the home page, and search page.
Solution in brief
Learn one shared set of customer and product embeddings, with a content-based network for new products. Each surface ranks its own candidate set by similarity to the customer vector, so a new surface reuses the same embeddings.
Questions this case answers
- How to rank items for an individual user on several surfaces at once while keeping inference latency low and the cost of adding a surface small
- How to rank newly added items that have no interaction history inside a collaborative recommender
- How to personalize product ranking across category, search, product and home pages with one model family?
- How to recommend newly added catalogue items that have no interaction history?
Tools and methods
- Neural collaborative filtering — turns interactions into product embeddings.
- Content-based product network — builds embeddings for new items from text and images.
- Customer embedding by aggregation — aggregates the vectors of products the customer interacted with.
- TensorFlow 2, Azure Machine Learning — training pipelines.
- TensorFlow Serving, NVIDIA Triton — serving the ranking models.
Google or a one-time email code. No password.