Problem
Pinterest's home feed was a grid of single Pins. A grid gives little context and no room for new topics, so the team added modules: carousels of Pins and entry points to landing pages. Every module displaces Pins in a feed that is already heavily optimized. Early on, launching new modules cost core company metrics, so launching new modules without a hit to core metrics was hard. Modules also vary in UI and content: topics, boards, recent engagement, shopping interests.
Example from the article: a shopping module shows clothing; a user may like its shoes and ignore its hats, and a module-level rule cannot tell the two apart.
Solution in brief
Three layers decide what a user sees. Fatiguing hides a module from users who keep ignoring it. A dedicated module ranker scores the remaining modules. Skip-slot blending puts a module in the feed only where its predicted engagement beats the Pin it would replace.
Questions this case answers
- How to blend a new content format into a feed that is already tuned for single items, so the new format only displaces an item when it is expected to engage more
- How to personalize which heterogeneous blocks a user sees when there is no engagement history for them yet and only simple signals are available
- How to add new content modules to a feed of items without losing core engagement?
- How to decide which modules to show a user and in what order?
Tools and methods
- Fatiguing — daily rule that hides a module from a user after many impressions without a click
- Pin Ranking model — existing ranker; baseline scorer for modules through their cover Pins
- Module ranker — multi-head model trained on module engagement, with a calibration layer
- Utility score — weighted sum of calibrated action predictions
- Skip-slot blending — comparison of module and Pin utility, slot by slot
Google or a one-time email code. No password.