Building a Unified Viewing Experience with Multi-Provider TV Recommendation
summary
An aggregator TV recommender that unifies on‑demand, partner catalogs and ~100 linear channels is built by consolidating messy multilingual metadata into a single content layer, applying matrix factorization for 20% of rating users, using implicit feedback (viewing frequency) for the rest, and leveraging LLMs to clean metadata and match duplicates achieving about 77% genre matching and shrinking effective item space by 35×, yielding a single collaborative model in production.
AI-written summary · about
chapters (12)
- 0:00Personalized recommendations intro
- 3:04Why standard methods failed
- 9:10Multi-provider challenges
- 13:30Algorithm approach
- 18:52Data consolidation approach
- 22:30LLMs for data cleaning
- 27:37Unified viewing from messy data
- 29:05Implicit feedback for TV data
- 31:58Recommendation engine end-to-end
- 33:50Data refinement over quick deployment
- 35:33Hyper-personalized trends
- 38:51Q&A
related talks
24mImproving YouTube Recommender systems for big screen users
19mX1 Recommendations Platform: Started from the bottom..
23mML-Powered Search & Recommendation 101
32mLet's talk about Recommendation Systems!
37mA primer on recommendation systems.
57mHands-On with LLM-Powered Recommenders: Hybrid Architectures
01h 21mUnlocking Personalization: A Deep Dive into Modern Recommendation Algorithms
25mBuilding a recommendation system for EPAM systems