Shopping Recommendations
summary
Pinterest turns a query product into a few hundred candidates from a 100M+ catalog by computing product embeddings with a graph convolutional network over the pin‑board graph plus image/text features and doing nearest‑neighbor lookup, then scores candidates with a multi‑head DNN that predicts clicks, long clicks and conversions, yielding single‑digit to double‑digit lift in engagement; page composition uses a multi‑armed bandit to choose module, image and title and gives about a 10% engagement gain.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:46How people use Pinterest
- 4:33Shopping on Pinterest
- 7:08Overview
- 9:34Shop Similar Recommendation
- 11:20How to Create Product Embedding
- 12:24Pin Board Graph
- 14:10Ranker
- 16:03Whole page optimization
- 19:25Summary
related talks
17mCalculating recommendations based on product images #bbuzz
21mUnleashing the Power of Data Science
24mMukund Narasimhan, Engineer, Pinterest at MLconf Seattle 2017
27mTransform online shopping experiences with personalized recommendations
23mML-Powered Search & Recommendation 101
22mHarnessing image and text data for product matching
39mAI-Powered Search in Modern E-Commerce
23mDeep Recommendations for e-Commerce