Recommender System for Playlist Continuation

Abstract

This project aims to develop and test different inputs for a recommender system for the task of automatic playlist continuation. The recommender system is built on a dataset of Spotify playlists created by users. The system attempts two main components: (1) a track similarity matrix, and (2) a k-means algorithm clustered on track metadata. With a partially filled playlist, the engine references these two components to deliver tracks the user is likely to enjoy. Track similarity matrix recommendations were able to achieve accuracy of up to 60% in sample, but at most 1% out of sample. K-means clustering was too computational expensive a task to provide any consequential recommendations.