Recsys Keynote: Improving Recommendation Systems & Search in the Age of LLMs
summary
Recommendation builds multimodal semantic IDs (Quaisho) that lift cold‑start coverage by 3.6% and raise cold‑start velocity, uses LLMs to synthesize labels (Indeed, Spotify) where GB4 yields ~90% precision/recall but is too slow, fine‑tuned GB3.5/GB2.5 then distills a lightweight classifier to <200 ms, cutting bad job recommendations ~20% while application rate rises 4% and unsubscribe falls 5%; Spotify gains +9% exploratory queries; unified models (Netflix, Etsy) match or beat specialists and drive +2.6% site conversion and >5% search purchases.
AI-written summary · about
chapters (12)
- 0:00Introduction to Language Modeling in Recommendation Systems
- 1:31Challenge 1: Hash-based Item IDs
- 2:14Solution: Semantic IDs
- 5:37Challenge 2: Data Augmentation and Quality
- 6:10Solution: LLM-Augmented Synthetic Data
- 6:21Indeed Case Study
- 10:37Spotify Case Study
- 13:34Challenge 3: Separate Systems and High Operational Costs
- 14:24Solution: Unified Models
- 14:51Netflix Case Study (Unicorn)
- 16:46Etsy Case Study (Unified Embeddings)
- 20:26Key Takeaways
related talks
23mML-Powered Search & Recommendation 101
30mLLM-Based Recommendation Systems
43mRecommendation Systems in Market Research
29mBuilding Generative Content Understanding Systems
41mRecommender Systems: Crafting Effective Personalised Experiences
57mHands-On with LLM-Powered Recommenders: Hybrid Architectures
37mSearch and Recommendations: 3 Sides of the Same Coin #bbuzz
20mBeyond Relevance: Using LLMs to Transform E-commerce Search into Product Discovery Engines