Contextual Set-Based Music Recommendations Using Interlocked Hypergraph Convolutional Neural Networks

Information Systems and Operations Management (ISOM) department

Speaker: Yang LI  (Cheung Kong Graduate School of Business - CKGSB)

Room Bernard Ramanantsoa

Abstract

"Contextual set-based personalized recommendations are important in many settings. For example, in the music industry playlists (i.e., sets of songs) are now the principal mechanism for music listening on digital streaming platforms. Given the large heterogeneity in musical tastes and the endless number of playlists that can be constructed, the automated design, completion, and recommendation of personalized playlists is critical. We develop a novel deep generative modeling framework to perform these tasks. We use interlocked Variational Hypergraph Auto-Encoders to uncover latent variables that summarize the musical themes and contextual moods associated with songs and playlists and the preferences of users. We fit our model on a data set of Spotify users’ playlists that we augmented with acoustic features and textual tags for the songs. Our application yields interesting insights about the diversity in musical preferences across users and contexts. We then use our model estimates to perform several managerial tasks such as similarity-based recommendations, complete existing playlists with congruent songs, and automatically design new playlists for personalized musical experiences."