Federated Clustering with Unknown Number of Clusters
Rong Zou, Yunfan Zhang, Yiqun Zhang, Yang Lu, Mengke Li, Yiu‐ming Cheung · 2024
Federated clustering is crucial to mining knowledge from unlabeled data distributed to multiple clients while preserving privacy. As there is no explicit learning supervision, clustering is considered a challenging federated learning task. Most existing works assume that the ‘true’ cluster number$k^{*}$is given to each client and server, which is far from a real federated learning scenario. Without the guidance of$k^{*}$, federated clustering becomes more challenging, rendering most existing solutions infeasible. We therefore propose a Federated Competitive and Cooperative Learning mechanism (FedCCL) to explore and fuse heterogeneous cluster distributions from clients automatically, and eventually form a global cluster partition, without requiring the cluster number to be given. We let the clients download seed points to explore their local distributions, which are then uploaded to the server for fusion. Different clients are allowed to compete on a single seed to form a consensus, while close seeds cooperate to represent a cluster. By iteratively homogenizing the cooperated seeds, a proper number of clusters will gradually emerge. Extensive experiments demonstrate the effectiveness of the proposed method.