Deep Multi-Objective Multi-Stakeholder Music Recommendation

Abstract

Many successful recommendation approaches rely on the optimization of a single objective function for a single stakeholder (i.e., the customer) based on customer-product interaction records. In this paper, we argue that for two-sided marketplaces such as music streaming platforms, it is more beneficial to consider multiple objectives for multiple stakeholders in the utility function for recommendations. We propose a novel recommender system design that aims at satisfying multiple---potentially conflicting---objectives of consumers and vendors. Specifically, we propose a "two-tower" neural network model for music recommendation that (a) learns each stakeholder objective in a different tower, (b) shares the latent information that was learned in each tower between the middle layers to predict each stakeholder objective, and (c) aggregates the predicted objectives using optimization methods to generate rating-based recommendations. In our model, we focus on user satisfaction objectives, such as saves, likes, and degrees of engagement with songs; and artist satisfaction objectives, such as acquiring new fans. We examine the performance of our proposed model on two proprietary industrial-scale datasets provided by Spotify and one public music listening history dataset from Last.fm. We find that our model is effective in solving the multi-objective problem and achieves the best performance trade-off for the two stakeholders (users and artists), when compared to the original single-objective model and state-of-the-art multi-objective multi-stakeholder methods.

Keywords: recommendation, neural networks, multi-objective, multi-stakeholder, transfer learning