Diversification in Recommender Systems
summary
Project creates new pin embeddings via matrix factorization (sampling boards with reservoir sampling), computes an embedding-based diversity score checked for stability, sensitivity and sanity, and applies it at the blender stage as a first-page optimization. Experiment shows users return more often, scroll deeper and view more impressions, with roughly a 1% increase in time spent, so the embedding-diversity intervention produces measurable, positive engagement lift.
AI-written summary · about
chapters (13)
- 0:00Introduction to talk
- 2:36Speaker background
- 5:25Motivation for diversification
- 7:09Diversification improves engagement
- 10:39Measuring diversity in recommender systems
- 16:59Sanity and sensitivity testing
- 20:11Pin embeddings for diversity
- 25:46Pinterest recommendation stack
- 27:28Implementing diversification
- 35:29First page optimization
- 38:17User studies and qualitative feedback
- 40:41Measuring experimental results
- 43:30Key takeaways and conclusion
related talks
32mLet's talk about Recommendation Systems!
37mA primer on recommendation systems.
21mUnleashing the Power of Data Science
41mRecommender Systems: Crafting Effective Personalised Experiences
01h 31mHands on - Build a Recommender system
35mHybrid Recommender Systems in Python
58mRecommender Systems – A Walkthrough
32mLeveraging recommender systems to personalise search results