Music Universe
An interactive music exploration system that organizes large-scale song data into a visual universe, helping users inspect genre structure and how musical styles evolve over time across the Million Song Dataset.
Overview
The project starts from a genre-labeled dataset and transfers that knowledge to a much larger song collection. After supervised genre classification, dimensionality reduction is used to map songs into an explorable spatial structure where each point represents one track.
The core target corpus is the Million Song Dataset, which provides large-scale audio feature data for roughly one million tracks. A labeled CD2C subset is used to train the genre model, then the predictions and reduced features are projected into a universe-style interactive view.
Video
Key Work
- Trained a supervised genre classification pipeline using labeled songs from the CD2C dataset.
- Used XGBoost to model hierarchical relationships between genres and predict labels for around one million songs.
- Applied PCA and UMAP to reduce high-dimensional audio features into a visual 3D structure.
- Built a universe-style interface where songs cluster into genre constellations and can be searched or explored spatially.
- Computed yearly genre centroids to trace how musical styles shift from 1955 to 2010.
