Bandit based Optimization of Multiple Objectives on a Music Streaming Platform
Recommender systems powering online multi-stakeholder platforms often face the challenge of jointly optimizing for multiple objectives, in an attempt to efficiently match suppliers and consumers. Examples of such objectives include user behavioral metrics (e.g. clicks, streams, dwell time, etc), supplier exposure objectives (e.g. diversity) and platform centric objectives (e.g. promotions). Jointly optimizing multiple metrics in online recommender systems remains a challenging task. Recent work has demonstrated the prowess of contextual bandits in powering recommendation systems to serve recommendation of interest to users. This paper aims at extending contextual bandits to multi-objective setting so as to power recommendations in a multi-stakeholder platforms. Specifically, in a contextual bandit setting, we learn a recommendation policy that can optimize multiple objectives simultaneously in a fair way. This multi-objective online optimization problem is formalized by using the Generalized Gini index (GGI) aggregation function, which combines and balances multiple objectives together. We propose an online gradient ascent learning algorithm to maximise the long-term vectorial rewards for different objectives scalarised using the GGI function. Through extensive experiments on simulated data and large scale music recommendation data from Spotify, a streaming platform, we show that the proposed algorithm learns a superior policy among the disparate objectives compared with other state-of-the-art approaches.