Building an Incrementally Trained, Local Taste Aware, Global Deep Learned Recommender System Model
summary
Netflix scale (about 150M accounts) drives a global recommender built as a variational autoencoder that encodes country, device and time by concatenating features or latent crosses, and masks out titles not available per region with a censored mask so parameters don’t waste capacity on absent catalog. New titles get randomly initialized embeddings and added to the softmax, then updated via warm-start incremental fine-tuning on top of the large global model; this pipeline retrains full models every few days and increments every few hours.
AI-written summary · about
chapters (19)
- 0:00Intro
- 0:10Netflix
- 0:46Recommendation Goals
- 1:08Personalization
- 2:23Collaborative Filtering
- 4:02Topic Models
- 5:11Global Model
- 5:29Catalog Differences
- 6:42Model Complexity
- 7:16Deep Learning
- 7:34Time Efficient
- 7:55Variational Auto Encoder
- 10:39Contextual Modeling
- 12:25The Problem
- 13:53Incremental Training
- 14:57Recipe
- 15:45Global Models
- 16:19Global Communities
- 17:30Conclusion
related talks
41mRecommendations for Building Machine Learning Software
42mDeep Learning Recommender Architecture: Master Thesis to ...
01h 31mHands on - Build a Recommender system
30mDeep Learning for Recommender Systems
29mFurther enhancements of recommender systems using deep learning
12mQuickUpdate: a Real-Time Personalization System for Large-Scale Recommendation Models
01h 21mUnlocking Personalization: A Deep Dive into Modern Recommendation Algorithms
32mLet's talk about Recommendation Systems!