Personalized Content Blending In The Pinterest Homefeed
summary
Pinterest builds a homefeed by ranking separate content pools (followed users, followed topics, recommendations) and blends them with a personalized integer ratio driven by a Thompson sampling bandit. Pin action probabilities are modeled as Beta-distributed pin action rates per action type, combined with action-specific reward weights to produce an expected utility per pool. The system maps utilities to small integer slot ratios, handles cold starts via priors, and increases positive-action rate and personalization in production.
AI-written summary · about
chapters (13)
- 0:00Intro
- 1:17Followed topics Danish Modern
- 1:44Followed users
- 1:57Recommendations
- 2:28How does the home feed work?
- 5:18The blending problem Goal Combine the results from the pools into one sorted feed.
- 7:23Which model?
- 9:10Model: Multi-armed bandit Each arm of the bandit represents one of the pools.
- 11:13Response distribution We should represent each type of action with its own beta distribution, and its own reward.
- 13:03Sampling technique
- 14:25Example
- 14:46User with history
- 17:21Personalizing the ratio
related talks
21mShopping Recommendations
03mServing Billions of Personalized News Feeds with AI
24mMukund Narasimhan, Engineer, Pinterest at MLconf Seattle 2017
46mDiversification in Recommender Systems
27mHow we personalized onet.pl with multi-armed bandits
21mRise of the Machines: Removing the Human-in-the-Loop
21mUnleashing the Power of Data Science
32mMoving the boundaries of possible at Pinterest