Recommender systems play a critical role in enhancing user experience on online platforms. Traditional recommendation models typically represent each user with a single embedding, but this potentially limits the capacity to reflect their diverse prefe...
Recommender systems play a critical role in enhancing user experience on online platforms. Traditional recommendation models typically represent each user with a single embedding, but this potentially limits the capacity to reflect their diverse preferences. Multi-interest recommendation (MIR) addresses this by representing users with multiple interest embeddings, each corresponding to a distinct interest. However, since each interest is still treated as a fixed point in the latent space, they capture only the central tendency of each interest, failing to reflect its detailed structure. To fully capture the dispersion within each interest, we propose the probabilistic multi-interest recommender, PRISM, representing each user as a mixture of multiple interest distributions. Each interest is modeled not just with a point embedding but also with its covariance, capturing both central tendency and spread of the preference. The final user representation is then composed by aggregating these multiple interests weighted by their relative importance. This distributional view allows a flexible and nuanced representation of user preferences, effectively modeling both their diversity and fine-grained structure. To further improve user-item matching, we introduce a distribution-aware scoring strategy that jointly considers the overall proximity and the interest spread. With extensive experiments, we verify that our method consistently outperforms existing MIR models at a negligible cost.