On a Framework for Federated Cluster Analysis

Abstract

Federated learning is becoming increasingly popular to enable automated learning in distributed networks of autonomous partners without sharing raw data. Many works focus on supervised learning, while the area of federated unsupervised learning, similar to federated clustering, is still less explored. In this paper, we introduce a federated clustering framework that solves three challenges: determine the number of global clusters in a federated dataset, obtain a partition of the data via a federated fuzzy c-means algorithm, and validate the clustering through a federated fuzzy Davies–Bouldin index. The complete framework is evaluated through numerical experiments on artificial and real-world datasets. The observed results are promising, as in most cases the federated clustering framework’s results are consistent with its nonfederated equivalent. Moreover, we embed an alternative federated fuzzy c-means formulation into our framework and observe that our formulation is more reliable in case the data are noni.i.d., while the performance is on par in the i.i.d. case.

Keywords

• federated learning
• framework
• cluster analysis
• cluster number determination
• federated fuzzy Davies–Bouldin index
• federated cluster validation metric