Article Details

Abstract

To address the limitation that existing federated recommendation methods predominantly rely on static hard gating strategies, which fail to capture the complex multifaceted nature and dynamic evolution of user interests, this paper proposes FedMoE, a dynamic Mixture-of-Experts (MoE) network model for federated recommendation. This is achieved by reconstructing a dynamic MoE architecture on the Federated Learning client. Firstly, a sparse soft routing strategy based on a Top-k retention mechanism is designed to supersede traditional hard clustering, thereby achieving fine-grained personalized modeling. Secondly, temporal context features are explicitly integrated into the gating network, empowering the model to perceive and adapt to user interest drifts across different time slices. Experimental results demonstrate that FedMoE, when integrated with various mainstream backbone networks, significantly outperforms existing state-of-the-art methods in terms of AUC, while maintaining highly competitive LogLoss performance.

Keywords

Federated Learning; recommender system; MoE network; temporal awareness

References

1. [1] Zhu Zhitao, Si Shijing, Wang Jianzong, etc Overview of Federated Recommendation Systems [J]. Big Data, 2022, 8 (4): 105-132
2. [2] Wang Yonghao, Xiao Feng, Wan Hongyou Overview of Personalized Federated Learning Algorithms [J]. Computer Technology and Development, 2026, 36 (2): 1-9
3. [3] Zhang Honglei, Li Wudong, Wu Jun, etc A review of privacy preserving federated recommendation algorithms [J]. Journal of Automation, 2022, 48 (9): 2142-2163
4. [4] ZHANG C X，LONG G D，GUO H K，et al. Multifaceted User Modeling in Recommendation: A Federated Foundation Models Approach [J/OL].arXiv:2412.16969 [cs.IR].[2025-12-28].https://arxiv.org/abs/2412.16969.
5. [5] ZHANG C，LONG G，GUO H，et al. Federated Adaptation for Foundation Model-based Recommendations [J/OL].arXiv:2405.04840 [cs.IR].[2025-12-28].https://arxiv.org/abs/2405.04840.
6. [6] TOUVRON H，MARTIN L，STONE K，et al. Llama 2: Open foundation and fine-tuned chat models [J/OL].arXiv:2307.09288 [cs.CL].[2026-1-08].https://arxiv.org/abs/2307.09288.
7. [7] GAO C M，LI S J，ZHANG Y，et al. Kuairand: An Unbiased Sequential Recommendation Dataset with Randomly Exposed Videos [C]//Proceedings of the 31st ACM International Conference on Information & Knowledge Management.ACM Digital Library，2022：3953-3957.
8. [8] SUN Z X，SI Z H，ZANG X X，et al. KuaiSAR: A Unified Search And Recommendation Dataset [C]//Proceedings of the 32nd ACM International Conference on Information and Knowledge Management.Birmingham：2023：5407-5411.
9. [9] PERIFANIS V，EFRAIMIDIS P S. Federated Neural Collaborative Filtering [J].Knowledge-Based Systems，2022，242：1-16.
10. [10] WANG C K，MCAULEY J. Self-Attentive Sequential Recommendation [C]//2018 IEEE International Conference on Data Mining.Singapore：IEEE，2018：197-206.
11. [11] ZHAI J Q，LIAO L，LIU X，et al. Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations [DB/OL].[2026-1-08].https://arxiv.org/pdf/2402.17152.

Funding

Graduate Education Innovation Plan Fund Project of Zhengzhou University of Aeronautical Industry Management (2025CX118)

About the authors

Xia Aoxiang (1999-), male, Han ethnicity, from Kaifeng, Henan, currently pursuing a master's degree, with a research focus on federated learning; Corresponding author: He Yuantao (1983-), male, Han ethnicity, from Zhengzhou, Henan Province, laboratory technician, master's degree, research direction: machine learning; Zhang Lianle (2003-), male, Han ethnicity, from Xinyang, Henan Province, is currently pursuing a master's degree with a research focus on federated learning.