摘要
在联邦推荐系统中,各客户端能否获得满意的推荐效果不仅取决于自身参与模型训练的数据,也取决于进行联合建模时其他客户端提供的数据。然而出于对自身数据安全的保护且数据获取不易,各客户端倾向于尽可能少地提供自身数据,期望其他客户端提供更多的数据来获得较好的推荐效果。文章首先使用不完全信息博弈模型对客户端之间的交互行为进行分析,接着引入满足均衡概念对该模型进行解释,假定各客户端均有一个预期推荐效果,当所有客户端都达到预期推荐效果时,即该博弈达到均衡。该文提出一种基于均衡学习的迭代算法,客户端通过分析当前推荐效果动态调整本地模型训练的数据,最终使各客户端均达到满足状态。理论分析和实验仿真表明,所提算法可以使各客户端均达到满足均衡,完成收敛。
In a federated recommendation system,whether each client can obtain satisfactory recommendation results depends not only on its own data involved in model training but also on the data provided by other clients when conducting joint modeling.However,due to the protection of their own data security and the difficulty in data acquisition,each client tends to provide as little data as possible and expects other clients to provide more data to obtain better recommendation results.This paper first uses the incomplete information game model to analyze the interaction behavior between clients,and then introduces the satisfaction equilibrium to explain the model,assuming that each client has an expected recommendation effect,and the game reaches equilibrium when all clients achieve the expected recommendation effect.An iterative algorithm based on equilibrium learning is proposed,which enables the client to dynamically adjust the training data by analyzing the current recommendation effect so that each client finally reaches the satisfaction state.Theoretical analysis and experimental simulation show that the proposed algorithm can make each client reach the satisfaction equilibrium and complete convergence.
作者
苏洋
张浩
刘俊彤
SU Yang;ZHANG Hao;LIU Juntong(School of Medical Information, Wannan Medical College, Wuhu 241002, China)
出处
《合肥工业大学学报(自然科学版)》
CAS
北大核心
2022年第5期625-632,共8页
Journal of Hefei University of Technology:Natural Science
基金
国家自然科学基金资助项目(61503116)
皖南医学院中青年科研基金资助项目(WK202116,WK202017)。
关键词
联邦推荐系统
博弈论
均衡学习
激励机制
隐私保护
federated recommendation system
game theory
equilibrium learning
incentive mechanism
privacy protection