music2vec: A smart music recommendation and discovery
summary
Uses a million playlists and ~300k song IDs to train a Python implementation that maps songs, playlists, albums and users into a single vector space; combines LDA and word2vec/paragraph2vec style PV embeddings, trains in a few hours with configurable params, and returns nearest neighbors by cosine similarity. Demonstrates analogies (song A − song B + playlist X), a web demo and 2D visualizations; live demo has glitches but embeddings show clear genre and artist clusters for recommendation via music2vec.
AI-written summary · about
chapters (10)
- 0:00Introduction to speaker
- 1:33Music vectorization concept
- 2:53Recap of Word2Vec
- 4:37Benefits of vector space
- 7:24Music recommendation systems
- 10:15Data processing and modeling
- 14:08Sequence and similarity
- 17:59Live technical demo
- 32:00Visualization and results
- 35:11Q&A and final remarks
related talks
40mBuilding A Folk Music Recommendation System with LLMs
43mpackage2vec: getting to know PyPI packages with ML
41mMusic and Machine Learning
01h 38mBuilding Recommender Systems Using Python
50mA fun and absurd introduction to Vector Databases
34mBuilding a Large Music Recommender Leveraging AI, Deep Learning and Human Expertise
28m"Embeddings: How Computers Learned to Read"
20mSo tell me, what is your musical taste?