Abstract
Recommender systems enhance user experience by delivering personalized suggestions derived from users’ historical behavior. However, conventional approaches face challenges in processing large-scale data while simultaneously preserving user privacy and maintaining stable training. To address these issues, we propose Fed-MWAE, a novel federated variational autoencoder (VAE) framework for recommendation tasks. The framework incorporates a sparsely activated Mixture-of-Experts (MoE) module to model diverse user behavior patterns across expert subnetworks. A top-<inline-formula><tex-math notation="LaTeX"></tex-math><alternatives><mml:math><mml:mi>k</mml:mi></mml:math><inline-graphic xlink:href="xiao-ieq1-3649484.gif"/></alternatives></inline-formula> gating mechanism selectively aggregates expert outputs, thereby improving computational efficiency without compromising accuracy. Furthermore, Fed-MWAE employs VAEs to capture complex latent structures and replaces the conventional Kullback-Leibler (KL) divergence with the Wasserstein distance, enabling smoother optimization and more stable convergence. Training is conducted in a federated learning setting, where local clients perform on-device updates to safeguard data privacy, and the updates are aggregated using the Federated Averaging (FedAvg) algorithm to enhance scalability and communication efficiency. Extensive experiments on four public datasets demonstrate that Fed-MWAE consistently outperforms strong baselines, achieving improvements of 5.46% in NDCG, 0.66% in Recall@20, 4.85% in Recall@50, and a 2.99% reduction in loss. These results validate the effectiveness of Fed-MWAE in balancing accuracy, efficiency, stability, and privacy in federated recommender systems.