Federated Consistency- and Complementarity-aware Consensus-enhanced Recommendation

Yunqi Mi, Boyang Yan, Guoshuai Zhao, Jialie Shen, Xueming QianarXiv 2025 · DOI · arXiv

Introduction

Personalized Federated Recommendation (FedRec) has emerged as a vital paradigm for providing high-quality suggestions while preserving user privacy. In a typical FedRec setup, a central server maintains a global item embedding table—a "consensus"—that is updated via aggregated gradients from decentralized clients. However, this consensus often suffers from a phenomenon the authors term Consensus Degradation.

Due to the inherent statistical heterogeneity of users, local updates often contain conflicting signals that cancel each other out during aggregation. The resulting global model becomes a diluted average that fails to accurately represent any individual user's specific distribution. Furthermore, existing methods struggle to decouple global knowledge from local preferences, often leading to redundant information or semantic mismatches that hurt recommendation accuracy.

The Fed3CR Framework

To address these challenges, the authors propose Fed3CR (Federated Consistency- and Complementarity-aware Consensus-enhanced Recommendation). Fed3CR is designed to treat the global consensus not as a final recommendation model, but as a source of "raw material" that must be refined locally through two primary components.

Adaptive Consensus Enhancement (ACE)

The ACE module introduces a personalized "transfer matrix" for each client. Rather than adopting the global item embeddings directly, the client learns to adaptively enhance specific signals within the global consensus. This transformation allows the client to map global knowledge into a local semantic space that better aligns with the user's historical interactions. By acting as a personalized filter, ACE ensures that the global information is "translated" into the local context before being used for prediction.

Consistency- and Complementarity-aware Optimization (C2O)

The C2O strategy manages the interplay between the global view (refined via ACE) and the purely local personal view. It relies on two key principles:

  1. Ranking-based Consistency: Traditional methods often attempt to minimize the Euclidean distance between global and local embeddings. Fed3CR argues this is too rigid, as it forces local models to mirror global scales. Instead, C2O employs a ranking-based loss (similar to ListNet). This ensures that both the global and personal views agree on the relative preference order of items, which is the primary goal of recommendation, while allowing for local nuances in embedding values.
  2. Orthogonality for Complementarity: To prevent the personal view from becoming a redundant copy of the global view, Fed3CR imposes an orthogonality constraint. By forcing the two embedding spaces to be linearly independent, the model ensures that the personal view captures unique, non-redundant features that the global consensus might have missed.

Experimental Evaluation

The authors evaluated Fed3CR across four diverse real-world datasets to test robustness across different domains and data densities:

  • MovieLens-1M: A relatively dense dataset for movie recommendations.
  • LastFM-2K: A sparse dataset for music recommendations.
  • Douban-Book: A dataset for book recommendations.
  • QB-article: A high-sparsity dataset involving article interactions.

Performance Results

Fed3CR demonstrated significant performance gains over state-of-the-art federated baselines such as FedRAP and GPFedRec. On average, the framework achieved an 8.60% improvement in Hit Ratio (HR@10) and a 21.09% improvement in NDCG@10 (a measure of ranking quality). Remarkably, in some cases, Fed3CR even outperformed centralized models like NeuMF, proving that localized federated learning can be more precise than a "one-size-fits-all" central model.

Key Insights and Versatility

A major takeaway from the research is the effectiveness of "soft" constraints. By prioritizing relative ranking over absolute similarity, Fed3CR allows local models to benefit from global trends without being stifled by them. Furthermore, the use of orthogonality proved superior to simple distance-based separation, providing a stronger mathematical guarantee of feature diversity.

Fed3CR is also notable for its plug-and-play versatility. When integrated into existing frameworks like FedMF and FedNCF, it provided immediate performance boosts, suggesting that the consensus degradation problem is widespread and that Fed3CR offers a generalizable solution.

Limitations and Future Work

While Fed3CR sets a new state-of-the-art, it introduces increased architectural complexity. The ACE module's transfer matrix can increase communication overhead if the underlying model becomes too deep. Future research will likely focus on optimizing these transfer functions for resource-constrained edge devices and exploring how Fed3CR handles extreme "cold-start" scenarios where local data is extremely limited.

Summary updated

Abstract

Personalized federated recommendation system (FedRec) has gained significant attention for its ability to preserve privacy in delivering tailored recommendations. To alleviate the statistical heterogeneity challenges among clients and improve personalization, decoupling item embeddings into the server and client-specific views has become a promising way. Among them, the global item embedding table serves as a consensus representation that integrates and reflects the collective patterns across all clients. However, the inherent sparsity and high uniformity of interaction data from massive-scale clients results in degraded consensus and insufficient decoupling, reducing consensus's utility. To this end, we propose a \textbf{Fed}erated \textbf{C}onsistency- and \textbf{C}omplementarity-aware \textbf{C}onsensus-enhanced \textbf{R}ecommendation (Fed3CR) method for personalized FedRec. To improve the efficiency of the utilization of consensus, we propose an \textbf{A}daptive \textbf{C}onsensus \textbf{E}nhancement (ACE) strategy to learn the relationship between global and client-specific item embeddings. It enables the client to adaptively enhance specific information in the consensus, transforming it into a form that best suits itself. To improve the quality of decoupling, we propose a \textbf{C}onsistency- and \textbf{C}omplementarity-aware \textbf{O}ptimization (C2O) strategy, which helps to learn more effective and complementary representations. Notably, our proposed Fed3CR is a plug-and-play method, which can be integrated with other FedRec methods to improve its performance. Extensive experiments on four real-world datasets represent the superior performance of Fed3CR.
BibTeX
@misc{mi2025federated,
  author = {Yunqi Mi and Boyang Yan and Guoshuai Zhao and Jialie Shen and Xueming Qian},
  title = {{Federated Consistency- and Complementarity-aware Consensus-enhanced Recommendation}},
  year = {2025},
  eprint = {2509.22659},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.22659},
}
Figure