Heterogeneous Graph Neural Network Music Recommendation

Abstract

While personalized music recommendation has changed the way many users listen to music. Graph Neural Networks have also become a state-of-the-art machine learning practice for predicting recommendations. The LFM-1b is a data set contains a high
density of information to address the sparsity issues of similar data sets like the The Million Song Data set. Additionally, as the provided information of the data set can be represented as a heterogeneous graph, their is a lot of available opportunities
to evaluate the important connections that users have with their favorite tracks, albums, artists, and even genres. However, as the music recommendation system research community has witnessed the promising capabilities of graph neural networks,
and as the limitations of a not having a publicly available, large scale, high in density data set as been alleviated, the LFM-1b data set is underutilized amidst the Music Information Retrieval community for graph based machine learning research.
This thesis will dive deep into the specifics required to utilize the LFM-1b data set for heterogeneous graph neural network research. With a primary focus on providing an machine learning recommendation system implementation, an analysis on the
models’ capabilities to provide recommendations to users whilst understanding user listening preferences is to be evaluated. The contributions of this thesis will include a LFM-1b data set loading class for the Deep Graph Library framework in Pytorch,
as well as an implementation of several link prediction graph neural network models, to validate the LFM-1b data set’s applicability for music recommendation system machine learning research.