Representation Learning with Graph Autoencoders and Applications to Music Recommendation

This talk will showcase part of Guillaume Salha-Galvan’s Ph.D. research at École Polytechnique (2018-2022) in partnership with the French music streaming service Deezer. We will focus on Graph Autoencoders (GAE). GAE emerged as a powerful family of unsupervised node embedding methods, with promising results in various graph-based machine learning problems including link prediction and community detection. Nonetheless, at the beginning of this Ph.D. project, GAE models were encountering critical limitations, hindering their industrial adoption. For instance, they were facing scalability challenges, obstructing their application to large industrial graphs with millions of nodes and edges. Moreover, they were designed for undirected and static graphs, whereas real-world data can be directed and/or dynamic. During this talk, we will discuss advancements made throughout this Ph.D. to overcome these challenges, aiming to enhance the applicability of GAE models for industrial-level problems involving graph representations. We will also present real-world applications of the proposed solutions to graph-based music recommendation at Deezer.