Hybrid Recommender Systems in Python
summary
Claims a hybrid matrix‑factorization that embeds item and user metadata and therefore handles cold start, reducing parameters by sharing feature embeddings. Implements as LightFM in Python, trains with multicore Hogwild SGD, uses ranking losses like WARP and BPR and per‑parameter learning rates. On StackExchange data pure CF gets AUC ~0.43, a content model improves, the hybrid reaches ~0.71. Fits offline, dumps vectors to Postgres C extensions for fast serving.
AI-written summary · about
chapters (19)
- 0:00Hybrid recommender systems intro
- 1:03Why hybrid for startups
- 3:53Matrix factorization
- 6:20Collaborative filtering limits
- 8:28Content-based drawbacks
- 11:16LightFM hybrid model
- 14:04Metadata embeddings benefits
- 16:08Evaluating on Stack Exchange
- 19:56Hybrid model benefits summary
- 20:17Installing LightFM in Python
- 21:21Scaling LightFM with Hog Wild
- 22:03Learning to rank
- 24:56Per-parameter learning rates
- 26:20Deployment and integration
- 27:01Q&A: SVD++ differences
- 28:05Q&A: Python in web apps
- 29:30Q&A: Temporal effects
- 30:33Q&A: Testing and diversity
- 32:42Q&A: Real estate models
related talks
58mRecommender Systems – A Walkthrough
37mA primer on recommendation systems.
01h 38mBuilding Recommender Systems Using Python
32mLet's talk about Recommendation Systems!
38mTowards Pythonic Innovation in Recommender Systems
01h 14mBuilding a Recommendation System with a Hybrid Search From Scratch
43m“Which car fits my life?” - mobile.de’s approach to recommendations
01h 31mHands on - Build a Recommender system