Recommender systems with Tensorflow
summary
Shows how to express matrix factorisation and feature-augmented embeddings in Tensorflow while using Pandas and Numpy for data prep, trains on MovieLens style data (movieLens 1M mentioned) and uses TensorBoard embedding projector to inspect latent space. Demonstrates binary click loss with negative sampling (1 positive + ~10 negatives), compares shallow ALS-like factors to deep YouTube/DSSTNE architectures, and describes candidate precompute / nearest-neighbour search for serving rather than real-time Tensorflow serving.
AI-written summary · about
chapters (17)
- 0:00Recommender systems intro
- 2:27Recap fundamentals
- 6:04Matrix factorization limits
- 8:57Spotify’s click prediction
- 11:03Deep learning frameworks
- 13:32TensorFlow inference graph
- 14:38Automatic differentiation
- 16:01User/item metadata
- 17:27Multi-class classification
- 18:52YouTube neural network
- 21:44Recommendations as search
- 22:04Embeddings with TensorBoard
- 24:32Candidate ranking architectures
- 25:16Training and deployment
- 26:40Wide & Deep architectures
- 27:00Time and sequence models
- 28:12Q&A
related talks
29mFurther enhancements of recommender systems using deep learning
37mA primer on recommendation systems.
01h 21mUnlocking Personalization: A Deep Dive into Modern Recommendation Algorithms
57mHands-On with LLM-Powered Recommenders: Hybrid Architectures
46mBeyond Bag of Words A Practitioner’s Guide to Advanced NLP
01h 31mHands on - Build a Recommender system
33mNeural Networks for Recommender Systems
33mHow to build personalization models using Amazon SageMaker