Reproduce “PinnerSage: Multi-Modal User Embedding Framework for Recommendations at Pinterest” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Aditya Pal, Chantat Eksombatchai, Yitong Zhou, Bo Zhao, Charles Rosenberg, Jure Leskovec
PinnerSage is a multi-modal user embedding framework developed by Pinterest's research team that represents each user with multiple embeddings, each corresponding to a distinct interest cluster, addressing the limitations of single-embedding user representations. Deployed in production, it increased engagement volume on Pinterest's home feed by 4% and shopping product recommendations by 20% by capturing diverse user interests more accurately.
Key insights
- Ward's hierarchical clustering effectively determines the optimal number of interest clusters per user and ensures conceptual coherence.
- Using medoids (actual user-interacted pins) to represent clusters provides interpretability, robustness, and efficiency benefits over centroids.
- PinnerSage simultaneously improves both recommendation relevance and diversity, achieving a better trade-off than single-embedding methods by sampling specific numbers of medoids per user.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.