Since streaming has emerged as the predominant means of music consumption, platforms such as Spotify rely on recommender systems to help users navigate within their increasingly large song catalogues. Collaborative filtering (CF) methods, which rely on past user-item interactions, have historically been most successful. They do, however, have various limitations, like performing poorly among lesser-known items with little or no interaction data. Conversely, content-based models attempt to circumvent the data-sparsity issue by generating recommendations based on item content (song audio) alone, but have seen limited success. A growing paradigm, which is seeing significant success in other fields and may help address the task at hand, is graph-based ML. Graph neural networks (GNN) in particular, promise to learn from both the complex relationships within a user-interaction graph, as well as from content to generate hybrid recommendations.
In our work, we introduce graph-based ML to the field of music recommendation by applying a state-of-the-art GNN, PinSage, to public Spotify data. We train the implemented algorithm on a newly collected playlist-song membership graph and evaluate it against traditional CF, graph-based CF and content-based methods on a related-song prediction task, venturing beyond-accuracy in our evaluation.
Our experiments show the following. Our approach is among top performers and stands out as the most well rounded. Graph-based CF methods significantly outperform matrix-based CF approaches, suggesting that user interaction data may be better represented as a graph.
In the scope of our evaluation task, CF methods do not exhibit a performance drop in the long tail, where the hybrid approach does not offer an advantage.
Based on our results we conclude that, although requiring further research, graph-based ML is a promising future direction for music recommendation.
|
