United StatesMedia and entertainment2023 – ongoingTransAct transformer for personalized recommendations
Pinterest reported an 11% increase in Homefeed repins in an online TransAct experiment
Problem
Recommendations needed to respond to changing user interests without losing the cost advantages of representations computed in batches.
Approach
Pinterest researchers built TransAct to extract short-term preferences from real-time actions. They combined it with user representations computed offline. The model reached production in Homefeed, Related Pins, Notifications and Search after online experimentation.
Results
Takeaway
Pinterest combined recent activity with longer-term representations in its ranking system. Retraining twice a week kept engagement gains stable as user interests changed.
Categories
- Industry
- Media and entertainment
- Business function
- IT and R&D
- Innovation type
- Product and service innovation
- Adoption
- Built in-house
- Evidence
- Company's own claim
Sources
Source summaries
Pinterest researchers reported an 11% increase in Homefeed repin volume in an online experiment. They described production deployment across recommendations, notifications and search.
Full details are in the original sources.
