FAACL: Federated Adaptive Asymmetric Clustered Learning
Keywords: Federated Learning, Distributed Learning, Clustering
Abstract: Asymmetric clustering has remained an unexplored problem in Clustered Federated Learning (CFL), diverging from the traditional approach of forming independent, non-interacting clusters. Previous methodologies have been limited to either separating devices with different data quality into distinct clusters or merging all devices into a single cluster, both of which compromise either data utilization or model accuracy. We propose a new federated learning technique where some devices may contribute to the training of the models of other devices, but without enforcing reciprocity, leading to a form of asymmetric clustering. This is beneficial in a variety of situations including scenarios where it is desirable for a device with high quality data to help train the model of a device with low quality data, but not vice-versa. This method not only enhances data utilization across the devices, but also maintains the integrity of high-quality data. Through a rigorous theoretical analysis and empirical evaluations, we demonstrate that our approach can efficiently find high quality (asymmetric) clusterings for numerous devices, achieving competitive performance metrics on existing CFL benchmarks.
Primary Area: other topics in machine learning (i.e., none of the above)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5019
Loading