Henry Zha Field Notes

Herd or heard? Disentangling social influence and personal taste in music listening with a probabilistic topic model

Hongyi Zhaa, Jiapeng Liua, Miłosz Kadzińskib, Jinnan Huanga, Xiuwu Liaoa,c

  • a Center for Intelligent Decision-making and Machine Learning, School of Management, Xi’an Jiaotong University, Xi’an, 710049, Shaanxi, PR China
  • b Institute of Computing Science, Poznan University of Technology, Piotrowo 2, 60-965, Poznań, Poland
  • c Collaborative Innovation Center of China Pilot Reform Exploration and Assessment - Hubei Sub-Center, Hubei University of Economics, Wuhan, 430205, Hubei, PR China

Omega Version of Record Volume 144 Available online Version of Record Last modified Article 103612

10.1016/j.omega.2026.103612

Highlights

  • We disentangle personal taste from herd effects in music listening at scale.
  • Spike–slab switch assigns each play to popularity or interest within the model.
  • Popularity is learned as a latent factor, while variational inference scales training.
  • On “30Music”, our model tops strong baselines; ablations and cold start confirm gains.
  • Outputs expose interpretable herding scores, interest mixtures, and item popularity.

Abstract

In large-scale music platforms, choices reflect both individual preferences and popularity trends. Understanding the interplay between these forces is a core challenge for developing an interpretable listening prediction framework. We propose the Probabilistic Disentanglement Model (PDM), which, for each play, determines whether the choice is interest- or popularity-driven. Our model characterizes users’ general preferences using a mixture of latent interests while concurrently modeling herd behavior by treating track popularity as a generative factor. A sparsity-inducing prior is incorporated to effectively disentangle these two components within a user’s listening history. For speed and scale, we use coordinate-ascent variational inference, which enables practical training on large datasets. On the Last.fm “30Music” dataset, PDM shows statistically significant gains in predictive accuracy over strong baselines based on matrix factorization, topic modeling, and neural collaborative filtering. Results are stable across different numbers of topics, and ablations show that removing latent popularity or the user/experience module reduces accuracy. The method is highly interpretable, yielding user herding scores, user interest mixtures, and item popularity estimates. They help understand listening behavior and manage popularity effects. In summary, PDM offers a straightforward, scalable, and interpretable approach to separating taste from popularity, thereby enhancing prediction and delivering valuable insights for auditing and control.

Citation

Zha, H., Liu, J., Kadziński, M., Huang, J., & Liao, X. (2026). Herd or heard? Disentangling social influence and personal taste in music listening with a probabilistic topic model. Omega, 144, 103612. https://doi.org/10.1016/j.omega.2026.103612