Applying Multi-Armed Bandit algorithms for music recommendations at Spotify

Abstract

This study explores the application of multi-armed bandit algorithms in enhancing music recommendation systems, with a focus on Spotify. It delves into the Explore-Then-Commit (ETC), Upper Confidence Bound (UCB), and Thompson Sampling (TS) algorithms, evaluating their efficacy within the Spotify context. The primary objective is to determine which algorithm optimally balances exploration and exploitation to maximize user satisfaction and engagement. The research reveals that the ETC algorithm, with its rigid exploration and exploitation phases, incurs a notably higher regret value. This rigidity can lead to missed opportunities in identifying optimal choices and hinder adaptability to evolving user preferences. Conversely, the UCB and TS algorithms exhibit remarkable adaptability and a flexible balance between exploration and exploitation. This flexibility translates into more personalized and satisfactory user experiences in music recommendations. However, the selection of the most appropriate algorithm should be contingent on the size and characteristics of the specific user dataset, as well as the fine-tuning of algorithm parameters to align with user preferences and behaviors.