The problem
Streaming services keep listeners by suggesting songs like the ones they already love, and artists want to know what makes a song rate well. We asked whether genre affects rating, whether explicit songs rate higher, and which audio features predict a rating.
How I built it
Data.
SQL queries against MySQL to process 95,000+ historical Spotify tracks.
Statistics.
t-tests and z-tests at a 95% confidence level, and ANOVA to compare ratings across genres.
Models and dashboard.
Multiple linear regression for song-rating prediction, presented in a user-friendly dashboard.
Recommender.
k-means clustering, with k = 4 chosen by the elbow method, returns the five songs closest to any track.
Drill-down.
The dashboard goes from genre to subgenre to artist to song, with an insights page on how each feature moves the rating.
Results
95,000+
historical tracks processed.
98%
confidence in genre-based rating differences, by ANOVA.
0.57
adjusted R² for song-rating prediction.
In hindsight
A recommender built this way favours songs and artists that are already popular, so new music struggles to surface.
