Balanced Federated Clustering via Anchor-Guided Dual Label Learning

Xingyu Xue, Quanxue Gao

Keywords: Machine Learning, Federated Multi-view Clustering, Anchor Graph, Balance Regularization

Abstract: Although the $\ell_{2,q}$-norm has been widely used in robust feature extraction and sparse modeling, its potential in promoting clustering balance has long been overlooked. This paper theoretically reveals the inherent ability of the $\ell_{2,q}$-norm to encourage balanced clustering, and proposes a federated multi-view clustering framework that incorporates it as a balance-aware regularizer. While preserving data privacy, the framework employs an efficient optimization strategy to learn a single label matrix, from which both anchor and sample labels can be inferred. The anchor labels then guide sample clustering, leading to improved clustering performance and robustness.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 24402