一种面向联邦推荐的动态混合专家网络模型

打开文本图片集

中图分类号：TP391.3 文献标识码：A 文章编号：2096-4706（2026）06-0070-06

A Dynamic Mixture--Experts Network for Federated Recommendation

XIAAoxiang，HEYuantao，ZHANGLianle （ ， ， 45o046，）

Abstract：To address the limitation that existing federated recommendation methods predominantly rely on static hard gating strategies，which fail to capturethecomplex multifaceted nature and dynamic evolution user interests，this paper proposes FedMoE，a dynamic Mixture--Experts （MoE） network model for federated recommendation.This is achieved by reconstructinga dynamic MoEarchitecture ontheFederated Learning client.Firstly，a sparse st routing strategy based ona Top-kretentionmechanismisdesignedtosupersede traditionalhardclustering，therebyachieving fine-grainedpersonalized modeling.Secondly，temporalcontextfeaturesareexplicitlyintegrated intothegating network，empoweringthemodel to perceive andadapt touser interest driftsacrossdifferent time slices.Experimentalresults demonstrate thatFedMoE，when integratedwithvarious mainstream backbone networks，significantlyoutperforms existing state--the-artmethods interms AUC，while maintaining highly competitive LogLoss performance.

Keywords： Federated Learning; recommender system; MoE network; temporal awareness

0 引言

推荐系统已成为解决信息过载、提升用户体验的核心技术。（剩余8214字）