Where To Next? A Dynamic Model of User Preferences Francesco Sanna Passino∗ Lucas Maystre Dmitrii Moor Imperial College London Spotify Spotify
[email protected] [email protected] [email protected] Ashton Anderson† Mounia Lalmas University of Toronto Spotify
[email protected] [email protected] ABSTRACT 1 INTRODUCTION We consider the problem of predicting users’ preferences on on- Online platforms have transformed the way users access informa- line platforms. We build on recent findings suggesting that users’ tion, audio and video content, knowledge repositories, and much preferences change over time, and that helping users expand their more. Over three decades of research and practice have demon- horizons is important in ensuring that they stay engaged. Most strated that a) learning users’ preferences, and b) personalizing existing models of user preferences attempt to capture simulta- users’ experience to match these preferences is immensely valuable neous preferences: “Users who like 퐴 tend to like 퐵 as well”. In to increase engagement and satisfaction. To this end, recommender this paper, we argue that these models fail to anticipate changing systems have emerged as essential building blocks [3]. They help preferences. To overcome this issue, we seek to understand the users find their way through large collections of items and assist structure that underlies the evolution of user preferences. To this them in discovering new content. They typically build on user pref- end, we propose the Preference Transition Model (PTM), a dynamic erence models that exploit correlations across users’ preferences. model for user preferences towards classes of items. The model As an example within the music domain, if a user likes The Beatles, enables the estimation of transition probabilities between classes of that user might also like Simon & Garfunkel, because other users items over time, which can be used to estimate how users’ tastes are who listen to the former also listen to the latter.