Learning to Collaborate via Structures: Cluster-Guided Item Alignment for Federated Recommendation

A research team led by Yuchun Tu and Zhiwei Li has published a novel paper, "Learning to Collaborate via Structures: Cluster-Guided Item Alignment for Federated Recommendation," introducing the CGFedRec framework. This work addresses a core inefficiency in federated learning for recommendation systems (FedRec), where conventional methods require synchronizing massive, high-dimensional item embeddings between a central server and distributed clients to achieve collaboration. The researchers challenge the assumption that precise geometric alignment of these embeddings is necessary, arguing that establishing relative semantic relationships is more effective and efficient.

The proposed CGFedRec framework transforms this paradigm. Instead of transmitting full embeddings, the server acts as a "global structure discoverer," learning item clusters and distributing only the resulting compact cluster labels to clients. This explicitly cuts off the downstream flow of dense embedding data. Clients then use these labels as structural constraints, allowing item representations to vary locally to capture user personalization while maintaining global semantic consistency. Extensive experiments show the method significantly improves communication efficiency—reducing data transfer by orders of magnitude—while maintaining superior recommendation accuracy, paving the way for more scalable and practical privacy-preserving AI.