Deep Recommendations for e-Commerce
summary
Facebook and Instagram Shops launch in May 2020 and recommendations use user interaction events (clicks, likes, purchases) rather than explicit ratings, so the system retrieves thousands then ranks to top 5–8 candidates. Models build dense embeddings + MLP stacks from sparse and dense features, trained with SGD variants across machines. Key engineering bottlenecks are embedding lookup memory bandwidth, network traffic for embedding interactions, and GPU compute for ranking; solutions include distributed/online training, parameter-server vs decentralized state, and frequent model refresh to track behavior drift.
AI-written summary · about
chapters (7)
- 0:00Start
- 3:41Overview of What the Recommendation System
- 5:32Matrix Factorization
- 8:09Stochastic Gradient Descent
- 13:28User Dynamics
- 18:20Memory Bandwidth
- 19:17Challenges
related talks
38mEmbeddings! Embeddings everywhere!
29mThe Challenges of Building a Personalized Recommendation System in E-Commerce
39mAI-Powered Search in Modern E-Commerce
23mML-Powered Search & Recommendation 101
44mRobust AI Search Ranking for Radical C2C Marketplace Growth
21mShopping Recommendations
17mEnhancing Chat-Based Customer Recommendations With GraphRAG and Hybrid Search
28mThe Power of Personalized Customer Experiences Through AI Innovations