Random Fourier Feature Shared Latent Variable Models for User Preference Visualization and Analysis

TMLR Paper5241 Authors

Abstract:

Understanding user preferences plays a crucial role in domains where strategies are designed by domain experts, such as personalized recommendations, targeted marketing, and human-centered interface design.
However, many existing methods prioritize predictive accuracy over model transparency, limiting their use in settings that require interpretability.

To address this issue, we propose the \textbf{Random Fourier Feature Shared Latent Variable Model (RFSLVM)}, a probabilistic generative model based on Gaussian processes that enables interpretable analyses of user preferences.
RFSLVM jointly models two data modalities: real-valued item features and binary user ratings.
It learns a \textit{two-dimensional} \textbf{visualization space} that captures relationships among items and user ratings.
Additionally, it infers \textit{user-specific} \textbf{preference vectors} that are compact and continuous representations of generally nonlinear preference functions.
These vectors support tasks such as measuring user similarity and performing preference-based clustering, thereby facilitating downstream analysis and decision-making.

We evaluate RFSLVM on multiple real-world datasets and find that it performs competitively against baseline models, while maintaining interpretability.
In addition, we demonstrate the utility of the learned representations through qualitative analyses, including hierarchical clustering and the identification of latent preference patterns.
These findings suggest that RFSLVM offers a practical and interpretable approach to modeling user preferences in real-world applications.

Submission Length: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Jaakko_Peltonen1

Submission Number: 5241