Beyond Discover Weekly: Building a Content-Based Spotify Discovery Engine

Publication Date

Spring 2026

Presentation Length

Poster/Gallery presentation

College

College of Sciences & Mathematics

Department

Math and Computer Science, Department of

Student Level

Undergraduate

Faculty Mentor

Dr. Davis

Metadata/Fulltext

Metadata ONLY

Presentation Type

Poster

Summary

Finding new music on streaming platforms like Spotify can be very difficult. Spotify recommends music based on what other people who have similar music tastes listen to. This often recommends songs you have heard before but have not added to your library. Even when you get a song recommendation that you haven’t heard, there is a 50% chance that you will like it, in my experience. But what if we instead recommended songs based on the musical DNA of the song? By analyzing song features such as energy and mood, we can define a specific listener’s music taste and understand their mathematical song relationships. This will, hopefully, allow us to give better recommendations.   For the dataset, I am using a Spotify song collection from Kaggle. I am also going to use an API from RapidAPI to get the audio features for some of my personal songs on Spotify. My approach compares PCA and Neural Network Autoencoders for feature extraction, followed by K-Means clustering to group songs based on these reduced latent representations. I will also engineer features in the hope of getting better results.   To measure success, I am going to see which dimensionality reduction technique best groups similar songs. To help visualize this, I will create interactive charts like UMAP. I will then create an average centroid of all the music features I pulled from my personal songs. After that, I will pull songs that are most similar to the centroid to see if the recommendation has a high cosine similarity and if I like the song. I will create multiple centroids, each being a different genre that I like. This approach will help uncover new songs that I have not heard and recommend music based purely on the music itself.

This document is currently not available here.

Share

COinS