Lessons from building deep learning recommendation systems
summary
Covers Taboola’s experience migrating their recommendation algorithms to deep learning over two years, focusing on architectures that handle multiple input types (click history, text, images). Describes concrete challenges and solutions in feature engineering for sparse recommendation data, techniques for capturing feature interactions, and how modeling choices affected system engineering and research practices.
AI-written summary · about
chapters (11)
- 0:00Intro
- 1:40Problem Setup
- 3:35Taboola algorithms
- 6:44Limitations of Matrix Factorization
- 8:52Factorization Machines
- 15:07We ask a lot from out models
- 20:00Embed Categories
- 21:00Adding a new feature
- 23:06Small incremental changes
- 26:20Modularity & End-to-end Training
- 29:16Summary
related talks
40mRecommendation Systems without the fuss
30mDeep Learning for Recommender Systems
32mLet's talk about Recommendation Systems!
29mFurther enhancements of recommender systems using deep learning
42mDeep Learning Recommender Architecture: Master Thesis to ...
01h 21mUnlocking Personalization: A Deep Dive into Modern Recommendation Algorithms
55mBuilding a Recommendation Engine with Machine Learning Techniques
37mDeep Recommender Systems