Discogs Style Embedding

Visualizing relationships between music subgenres from Discogs.

Motivation

• Discogs categorizes releases and artists into 15 genres and 760 styles (subgenres).
• Many subgenres co-occur in the same release—similar to how words co-occur in sentences.
  • Example: Post-Rock and Shoegaze often appear together on the same album, showing that these genres are closely related.
• Can we capture these relationships in a vector space?

Approach

1. Represent 760 subgenres as low-dimensional vectors using:
   • Word2Vec (treating co-occurring styles like co-occurring words)
   • Node2Vec (treating styles as nodes in a co-occurrence graph)
2. Visualize embeddings with dimensionality reduction (UMAP) in 2D and 3D.

Map Visualizations

Method	2D	3D
Word2Vec	View	View
Node2Vec	View	View

Data Sources

• Discogs monthly data dump: https://data.discogs.com/
• Current version: 2026-01-01
• Note: Word2Vec.py and counter.py use raw_data/discogs_20260101_masters.xml.gz.

  This file is not included due to its large size.