大数据

• •    

联邦推荐系统综述

朱智韬1,2 , 司世景1 ,王健宗1,* , 肖京1   

  1. 1平安科技(深圳)有限公司, 深圳 518063
    2中国科学技术大学, 合肥 230026
  • 作者简介:朱智韬 (1996-),男,中国科学技术大学硕士研究生,平安科技(深圳)有限公司算法工程师,中国计算机学会会员,主要研究方向为人工智能,联邦学习和推荐系统等。 司世景(1988‒ ),男,博士,平安科技(深圳)有限公司资深算法研究员,深圳市海外高层次人才。美国杜克大学人工智能博士后,中国计算机学会会员,主要研究方向为机器学习和及其在人工智能领域应用。王健宗(1983‒ ),男,博士,平安科技(深圳)有限公司副总工程师,资深人工智能总监,联邦学习技术部总经理。美国佛罗里达大学人工智能博士后,中国计算机学会高级会员,中国计算机学会大数据专家委员会委员,曾任美国莱斯大学电子与计算机工程系研究员,主要研究方向为联邦学习和人工智能等。 肖京(1972‒ ),男,博士,中国平安集团首席科学家,2019年吴文俊人工智能杰出贡献奖获得者,中国计算机学会深圳分部副主席。主要研究方向为计算机图形学学科、自动驾驶、3D显示、医疗诊断、联邦学习等。

Survey on Federated Recommendation Systems

Zhu Zhitao1,2, Si Shijing1, Wang Jianzong1,*, Xiao Jing1   

  1. 1. Ping An Technology (Shenzhen) Co., Ltd., Shenzhen, 518063
    2. University of Science and Technology of China, Hefei, 230000

摘要: 自联邦学习被提出以来,针对其优化方法和部署实现的快速发展步伐从未放缓。 在联邦学习范式中,原始数据被本地存储在独立的用户客户端中,而脱敏数据被发送到中心服务器中加以聚合,这给众多领域提供了一种新颖的设计思路。考虑到传统推荐系统的研究方向集中于提高推荐效果,在资源节约、跨领域推荐、隐私保护等方面还具有很大改进空间,如何将联邦学习与推荐系统结合以缓解这些问题成为当前的一个研究热点。本文第一次对近年来基于联邦学习的推荐系统进行了全面的总结、比较与分析,首先介绍了推荐系统的传统实现方式及面临的瓶颈,然后,引入了联邦学习范式,并描述了联邦学习在隐私保护、利用多领域用户数据两方面给推荐系统带来的增益,以及二者结合的技术挑战,进而详细说明了现有的联邦推荐系统部署方式。 最后,对联邦推荐系统未来的研究进行了展望与总结。

关键词: 联邦学习, 推荐系统, 隐私保护, 协同过滤, 深度学习

Abstract: In the federated learning (FL) paradigm, the original data are stored in independent clients while masked data are sent to a central server to be aggregated, which proposes a novel design approach to numerous domains. So motivated, the fast pace of its optimization and deployment never decelerates after the proposal of FL. Given the wide application of recommendation systems (RS) in diverse domains, combining RS with FL techniques has been gaining momentum to reduce the computational cost, do cross-domain recommendation and protect users’ privacy while maintaining recommendation performance as traditional RS. To the best of our knowledge, the federated learning based recommendation systems in recent years were comprehensively summarized for the first time in this work. We analyzed the difference between traditional and federated recommendation systems, and demonstrate the main research direction and progress of federated recommendation systems with comparison and analysis. First, we summarize the traditional recommendation systems and their bottleneck. Then the federated learning paradigm is introduced. Furthermore, we depict the advantages of combining federated learning with recommendation systems in two aspects: privacy protection and usage of multi-domain user information, along with the technical challenges during the combination. At the same time, the existing deployment of federated recommendation systems is illustrated in detail. Finally, the future research of federated recommendation systems is prospected and summarized.

Key words: federated learning, recommendation system, privacy-preserving, collaborative filtering, deep learning

No Suggested Reading articles found!