Building a Large Music Recommender Leveraging AI, Deep Learning and Human Expertise
summary
Pandora combines human-curated Music Genome Project annotations with machine listening and deep learning models trained on data from ~70 million monthly users, ~13 billion personalized stations and 95+ billion thumbs to predict song attributes (measured by AUC) and power an ensemble of 80 algorithms. Models tag up to 450 attributes per song, recompute nightly, run online A/B and year-long holdouts, and pick the next track within 50-80 ms; deep learning improves accuracy but increases complexity and deployment cost.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:00Welcome
- 6:50What is Pandora
- 7:40Lessons Learned
- 7:55Big Data
- 9:35Model Selection
- 14:45Business Goals
- 15:30Offline Evaluation
- 16:18Online Evaluation
- 18:27Experimentation and Production
- 19:30Framework
- 24:30Domain Expertise
- 25:00Why is Music Different
- 27:24Dont do stupid things
- 28:32Summary
- 29:26QA
related talks
05mData science and streaming music - Interview with Jonathan Gosier
45mWorkshop: Building a Music Recommender System in Java
04mMusic, the window into your soul
42mDeep Learning Recommender Architecture: Master Thesis to ...
27mPlay
38mmusic2vec: A smart music recommendation and discovery
40mPowering Spotify's Audio Personalization Platform
28mThe Power of Personalized Customer Experiences Through AI Innovations