Keep It or Skip It? Sequential Consumption of Music with Reference Effects

Sequential consumption of experiential products gives rise to inter-temporal associations. Developing recommendation algorithms that account for these effects while designing experiences for users can be effective in enhancing user engagement. Using music streaming as the paradigmatic context of such interactions -- consumption of multiple songs across multiple sessions -- we construct a utility-based theoretical framework that accounts for users' past consumption, leading to: (a) Recall-based references, that are built on past sessions, and (b) Locally-based references, which are the result of previous songs in the focal session. Users' heterogeneous responses, rooted in the constructs of habit formation and variety seeking, help us understand their dynamic preferences, which are further influenced by the memory decay effects. To validate this theory, we combine music streaming logs of 44,794 paid customers of the global streaming platform, Deezer, across 2018-2019 with the song-attribute data from Spotify to study the relationship between the platform-recommended song-attribute deviation from user references and user engagement with the platform. Engagement is measured via song skipping, duration of song listened to, and session abandonment decisions. Using a matching procedure in conjunction with reduced-form analyses, we find that a 1% increase in the deviation of the recommended song's attributes from the Recall and Local references results in an increase of 0.2% (variety seeking) and a drop of -1.9% (habit formation) in the engagement. Finally, the counterfactual analyses show an increase in user engagement levels of 34.7% above the status-quo at Deezer when implementing our recommendations. An experimental study further supports these results