eFedGauss: A Federated Approach to Fuzzy Multivariate Gaussian Clustering
Miha Ožbot, Seiichi Ozawa, Igor Škrjanc · 2024
In this paper, we present a novel federated soft clustering method for unsupervised clustering of data and classification, termed Evolving Federated Gaussian clustering (eFedGauss). There is a growing need for Federated Clustering methodologies; however the main drawback of traditional clustering methods is the requirement to select the number of clusters a priori. This is a significant challenge in Federated Learning where data is commonly non-identically distributed. We propose solving this problem by using Evolving Fuzzy Clustering, which has mechanisms to dynamically add and remove clusters. Our proposed methodology was tested on unsupervised clustering using synthetic data, classification of the Iris Flower, and an unbalanced Credit Card Fraud Detection. The results demonstrate the benefits of federated clustering over the non-federated approach, outperform related evolving clustering methods on the iris dataset, and show promising results in high-dimensional credit fraud detection.