Music Vibes
Exploratory music / preference analytics under the Futuro Insights shell — clustering, embeddings, and playful signal design for audio taste.
A playlist is a soft clustering problem with a beat: you propose a distance on “vibe,” then check whether the next track still belongs in the room.
Where to listen and read
Sam Sounds
The music site is the listening room. This page is the analytics note beside it.
Why it lives here
Audio preference data is another lab for embeddings, k-means, and dataviz — cousins of the movie network and clustering notes.
Modeling angle
Treat tracks as points in feature space (tempo, energy, valence, genre tags). Segments become playlists; outliers become “skip” signals.
Working notes
- Features before fancy models. Start with hand-built descriptors and simple distances. If clusters don’t stabilize, more layers won’t save you.
- Cold start is the product. New users and rare tracks need priors — same credibility story actuaries tell about sparse cells.
- Evaluate like a human. Silhouette scores help; a listening test on held-out tracks tells you if the grammar of “similar” is honest.
- Downstream jobs. Use clusters as strata for recommendations, A/B treatments, or playlist generators — judge them by that job, not by inertia alone.
Listen on Sam Sounds. The source repository is SamSoundsVibes.