World's Fair 2025
Recsys Keynote: Improving Recommendation Systems & Search in the Age of LLMs - Eugene Yan, Amazon
Overview
This talk explores the integration of Large Language Models (LLMs) with recommendation systems and search functionalities. It proposes three key areas for advancement: semantic IDs to better represent item content and address cold-start problems, data augmentation using LLMs to improve data quality and scale for training, and unified models to streamline complex recommendation infrastructures.
Who should watch
- AI Engineers
- Product Managers
- Builders working on recommendation systems
- Individuals facing challenges with cold-start problems in recommendations
- Teams looking to improve search result quality and relevance
- Engineers seeking to simplify complex recommendation architectures
Key takeaways
- Traditional hash-based item IDs in recommendation systems do not encode item content, leading to cold-start and sparsity issues. Semantic IDs, potentially incorporating multimodal content, can address this by understanding item characteristics.
- LLMs can significantly enhance data augmentation for recommendation systems and search by generating synthetic data and labels, reducing the cost and effort compared to human annotation.
- Unified models, similar to those used in vision and language, can consolidate disparate recommendation and search systems into a single, more efficient architecture, reducing maintenance costs and improving cross-use case learning.
- The Quaou platform used trainable multimodal semantic IDs, combining visual, textual, and audio embeddings with clustering, to improve cold-start coverage and velocity for short videos.
- Indeed used a fine-tuned LLM and then distilled a lightweight classifier to filter bad job recommendations, resulting in a 20% reduction in poor suggestions and a 4% increase in application rates.
- Spotify leveraged LLMs to generate natural language queries, augmenting traditional methods to drive exploration of new content categories like podcasts and audiobooks, leading to a 9% increase in exploratory queries.
- Netflix developed a unified contextual ranker (UNICORN) that integrates user history and content context for various tasks like search and recommendations, matching or exceeding the performance of specialized models.
- Etsy implemented unified embeddings by sharing encoders for product and query data, incorporating a quality vector to improve search relevance and conversion rates by 2.6%.
Notable quotes
*The long story short, the benefits of semantic ids, you can address coart with the semantic ID itself and now your recommendations understand content.*
*The lifeblood of machine learning is data. good quality data at scale and this is very essential for search and of course recommendation systems.*
*So the solution for this is unified models, right? I mean it works for vision, it works for language. So why not recommendation systems?*
Unofficial community note. Prefer the recording for nuance.