- 分享主题：Cluster-driven Graph Federated Learning
- 论文标题：Cluster-driven Graph Federated Learning over Multiple Domains
- 论文链接：https://arxiv.org/pdf/2104.14628v1.pdf
- 分享人：唐共勇

1. Summary

【必写】，推荐使用 grammarly 检查语法问题，尽量参考论文 introduction 的写作方式。需要写出

  1. 这篇文章解决了什么问题？
  2. 作者使用了什么方法（不用太细节）来解决了这个问题？
  3. 你觉得你需要继续去研究哪些概念才会加深你对这篇文章的理解？

FedCG is the first method to model domain-domain interaction through GCN. In GCN, each node is composed of domain-specific model parameters, and the adjacency matrix is composed of the reverse distance between domain-specific parameters. Based on this, FedCG not only captures the particularity of each domain but also allows each domain to benefit from the update of other domains and share the knowledge of different domains during training. Clustering in FedCG is based on unsupervised teacher-student classifier training, and it is extended to the invisible test domain. Train according to the pseudo label assigned by the teacher and learned by students, and consider the problem of data privacy and security. Estimate the soft assignments of the invisible new test domain through the cluster. This paper explores the combination of FL and graph representation learning to solve the statistical heterogeneity of data distribution and uses graphs to learn the parameters of specific fields and model the interaction between them.

2. 你对于论文的思考

需要写出你自己对于论文的思考，例如优缺点，你的takeaways

  1. 通过符合FL标准的聚类确定领域，并为每个领域实例化特定领域模块；
    2. 在训练时通过 GCN连接特定领域的模块，以学习各领域之间的相互作用并分享知识；(等价于一个共享信息的选择器）
    3. 通过 teacher-student 分类器无监督学习 cluster，并通过其领域 soft-assignment 分数处理没有见过的测试领域。

    3. 其他

    【可选】

image.png
teacher给出label
student根据label训练修正
一定阶段后，用student参数更新teacher参数
集群特定模型：通过分类结果加权不同域对应模型组件
具体聚类过程：
假设原始数据包含 Cluster-driven Graph Federated Learning - 图2 个领域， Cluster-driven Graph Federated Learning - 图3 为超参，本文初始化两个分类器分别为教师分类器 Cluster-driven Graph Federated Learning - 图4 和学生分类器 Cluster-driven Graph Federated Learning - 图5 （参数分别为 Cluster-driven Graph Federated Learning - 图6 ）。本文应用的式 CV 领域，每一个域分类器是一个函数，映射图片到一个概率向量 Cluster-driven Graph Federated Learning - 图7 。教师分类器的预测结果作为 pseudo-label 充当学生分类器预测目标。在训练过程中通过最小化基于本地数据集 Cluster-driven Graph Federated Learning - 图8 的教师分类器 Cluster-driven Graph Federated Learning - 图9 和学生分类器 Cluster-driven Graph Federated Learning - 图10 的域预测结果来训练本地学生参数 Cluster-driven Graph Federated Learning - 图11 ：
Cluster-driven Graph Federated Learning - 图12
其中 Cluster-driven Graph Federated Learning - 图13 是由教师分类器基于 Cluster-driven Graph Federated Learning - 图14 给出的 pseudo-label，形式化表示为 Cluster-driven Graph Federated Learning - 图15 ，并且 Cluster-driven Graph Federated Learning - 图16 表示为基于分类器 Cluster-driven Graph Federated Learning - 图17 ，其中 Cluster-driven Graph Federated Learning - 图18 属于第 Cluster-driven Graph Federated Learning - 图19 个领域的概率。式（3）鼓励学生能够根据伪标签进行分类，并隐含地鼓励在伪标签上达成一致，从而在聚类上达成一致。并且在 FL 设定下，域分类器的参数在每一个 round 基于标准 FedAvg 进行更新： Cluster-driven Graph Federated Learning - 图20 。
本文的聚类算法可以基于域分类器在测试时将未见过的数据分配到聚类中。Cluster-driven Graph Federated Learning - 图21 是 soft assignment，可以通过现有领域的组合来适应属于未见过的领域的数据。