Community of Experts-Based VAE for Recommendation

User populations in modern recommendation systems consist of diverse communities with distinct and often non-overlapping preferences. However, most existing models, such as traditional Variational Autoencoders (VAEs), optimize for global performance and learn a single, averaged user representation. This monolithic approach overlooks underlying community structures and fails to capture the multimodal nature of user preferences, resulting in limited personalization and adaptability. To address this limitation, we propose the Community-of-Experts VAE (CoE-VAE), a deep generative model that explicitly discovers and models user communities through specialized expert sub-models. Our method introduces a similarity-guided training paradigm, where a Mixture-of-Experts (MoE) gating network is trained to reconstruct a cosine similarity matrix derived from users’ implicit interaction vectors. This mechanism provides a strong, data-driven signal that encourages expert specialization, allowing users to be dynamically assigned to semantically coherent expert communities. This similarity-guided MoE framework is jointly regularized by a Gaussian Mixture Model (GMM) prior that shapes the latent space to capture distinct user modalities. Extensive experiments demonstrate that CoE-VAE consistently outperforms strong generative baselines on all benchmark datasets. Furthermore, it achieves performance highly competitive with state-of-the-art Graph Neural Networks (LightGCN). It effectively mitigates expert collapse and learns well-defined, interpretable user communities, validating the effectiveness of our similarity-guided specialization mechanism for improving recommendation accuracy.