MusicDress: A Heterogeneous Dataset for Comparing Music Recommender Systems

Abstract

To compare different types of music recommender systems, datasets are necessary that offer a combination of diverse features. We propose MusicDress, a novel dataset covering four different elements of music: timbre, rhythm, melody, and harmony. The dataset extends to lyrics and user data by linking to publicly available data sources. It comprises features of 2,136 individual songs and enables the comparison of hybrid recommender systems that combine content-based, context-based, and collaborative filtering approaches.