Recent research has identified key weaknesses in variational autoencoder (VAE) models used for generative recommender systems, particularly their reliance on uniform priors that fail to capture diverse user behavior. This limitation leads to poor generalization, weak personalization, and stability issues like posterior collapse.
To overcome these challenges, a new study proposes an enhanced VAE model featuring a Dynamic Hierarchical Mixture Prior and a learnable temperature-scaling mechanism. The approach integrates global, user-specific, and uniform priors to better model heterogeneous user preferences while improving reconstruction stability.
The proposed method aims to deliver more accurate and personalized recommendations by adapting to individual user patterns, addressing a long-standing bottleneck in generative recommendation systems.