Abstract: In the cloud manufacturing service scenario,service composition developers often need to develop customized service compositions based on users' manufacturing needs.With the prevalence of privacy protection laws and regulations,it has been difficult to implement commonly used user demand mining algorithms such as LDA topic model in practice.Through the cross-study of blockchain and federated learning technologies for user demand mining,a Decentralized Federated Latent Dirichlet Allocation algorithm (DAFedLDA) was proposed.In DAFedLDA,a Multi-Channel Access Control Scheme (MCACS) and a Random-Dropout Data Monitor Scheme (RDDMS) were presented based on the peer-to-peer distributed LDA.On the basis of the ProgrammableWeb.com dataset,a series of experiments demonstrated the effectiveness of the proposed algorithm.

Key words: user demand mining, cloud manufacturing service, blockchain, federated learning, Latent Dirichlet allocation

摘要： 在云制造服务场景下,服务组合开发者往往需要基于用户的制造需求进行定制化服务组合开发。随着隐私保护法律法规的相继颁布,常用的用户需求挖掘算法如隐狄利克雷分布(LDA)主题模型已难以在实际中使用。本文通过对区块链和联邦学习技术交叉研究,提出了面向用户需求挖掘的去中心化异步联邦隐狄利克雷分布算法(DAFedLDA)。在DAFedLDA中,本文基于对等分布式LDA,进一步提出了基于多链的权限控制机制(MCACS)以及基于随机丢弃的数据贡献质量监控机制(RDDMS)。本文基于ProgrammableWeb.com实例进行了一系列实验,验证了算法的有效性。

关键词: 用户需求挖掘, 云制造服务, 区块链, 联邦学习, 隐狄利克雷分布