Explore a comprehensive discussion on recommender systems at Spotify in this 50-minute podcast episode featuring Sanket Gupta, Senior Machine Learning Engineer. Dive into the construction and maintenance of large-scale recommender systems, the use of foundational user and item embeddings for transfer learning across products, evaluation methods, and MLOps challenges. Learn about balancing long-term and short-term user understanding for personalized experiences, cold start problems, real-time recommendations, and the integration of vector databases. Gain insights on Spotify's product integration strategy, metrics for speed and relevance, and staying current with new features in the rapidly evolving field of music recommendation systems.
RecSys at Spotify - Building and Evaluating Large-Scale Recommender Systems
Overview
Syllabus
[] Sanket's preferred coffee
[] Takeaways
[] RecSys are RAGs
[] Evaluating RecSys parallel to RAGs
[] Music RecSys Optimization
[] Dealing with cold start problems
[] Quantity of models in the recommender systems
[] Radio models
[] Evaluation system
[] Infrastructure support
[] Transfer learning
[] Vector database features
[] Listening History Balance
[26:35 - ] LatticeFlow Ad
[] The beauty of embeddings
[] Shift to real-time recommendation
[] Vector Database Architecture Options
[] Embeddings drive personalized
[] Feature Stores vs Vector Databases
[] Spotify product integration strategy
[] Staying up to date with new features
[] Speed vs Relevance metrics
[] Wrap up
Taught by
MLOps.community
Related Courses
-
Building Recommender Systems with Machine Learning and AI
-
From Robotics to Recommender Systems - MLOps Podcast #240
-
Unpacking Three Types of Feature Stores - MLOps Podcast
-
Build a Spotify-Like Music Recommender System in Python
-
RAG Has Been Oversimplified - Exploring Complexities in Retrieval Augmented Generation
-
Adventures in Building CLIP and Other Large Language Models
