Research map / Federated learning and privacy / Personalized and heterogeneous FL
Asynchronous Federated Clustering with Unknown Number of Clusters
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AFCL, an asynchronous federated clustering method that spreads excess seed points across clients, accumulates client-side update intensities, and uses a balance-weighted server-side seed interaction to learn the number of clusters without knowing k*.
Problem
Federated clustering on non-IID data where clients upload asynchronously at heterogeneous rates and the true number of clusters k* is unknown.
Approach
Clients learn update intensities for an over-provisioned set of seed points from local data without updating them; the server fuses these intensities via seed-to-seed interaction (neighboring seeds within a radius share updates) with a communication-frequency-based balance weight, so redundant seeds merge and the number of clusters emerges.
Key results
- AFCL achieves the best average SC rank of 1.0769 across 13 datasets, versus 4.0000 for the next best (FedSC).
- On the AL dataset, AFCL reaches SC 0.6138 versus 0.4422 for the next best (Fed-SC), and CH 4880.6 versus 2260.6.
- AFCL achieves the best CH average rank of 1.5385 across 13 datasets, with the best CH on 8 of 13 datasets (per the reported tables).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AFCL instead of the strongest listed baseline | AFCL · SD1 | SC: 0.8539±0.00 -> 0.9714±0.00 | Helped |
| Use AFCL instead of the strongest listed baseline | AFCL · SD2 | SC: 0.7699±0.00 -> 0.8571±0.00 | Helped |
| Use AFCL instead of the strongest listed baseline | AFCL · SE | SC: 0.4323±0.05 -> 0.5033±0.09 | Helped |
| Use AFCL instead of the strongest listed baseline | AFCL · IR | SC: 0.6119±0.21 -> 0.6386±0.06 | Helped |
| Use AFCL instead of the strongest listed baseline | AFCL · AL | SC: 0.4422±0.05 -> 0.6138±0.38 | Helped |
| Use AFCL instead of the strongest listed baseline | AFCL · AB | SC: 0.4863±0.26 -> 0.5005±0.11 | Helped |
| Use AFCL instead of the strongest listed baseline | AFCL · CC | SC: 0.3809±0.02 -> 0.5916±0.04 | Helped |
| Use AFCL instead of the strongest listed baseline | AFCL · AC | SC: 0.2656±0.13 -> 0.4851±0.26 | Helped |
| Use AFCL instead of the strongest listed baseline | AFCL · SG | SC: 0.3865±0.10 -> 0.6086±0.12 | Helped |
| Use AFCL instead of the strongest listed baseline | AFCL · LI | SC: 0.8256±0.01 -> 0.8967±0.01 | Helped |
| Use AFCL instead of the strongest listed baseline | AFCL · PA | SC: 0.5138±0.00 -> 0.4903±0.11 | Hurt |
| Use AFCL instead of the strongest listed baseline | AFCL · AU | SC: 0.4296±0.00 -> 0.5191±0.07 | Helped |
Surprising findings
- The objective function does not decrease monotonically because only limited statistics are shared between clients and server, so strict gradient descent is not guaranteed.
Limitations
- Assumes federated clustering on purely numerical data.
- Evaluated with a small number of clients (five clients simulated by k-means partitioning).
- Privacy protection is not the focus; differential privacy or homomorphic encryption are only suggested as add-ons.
Experimental setup
Datasets: SD1 SD2 Seeds Iris Avila Abalone Breast Cancer Accent Segment Live Parkinson Audit
Benchmarks: SD1 SD2 Seeds Iris Avila Abalone Breast Cancer Accent Segment Live Parkinson Audit
Metrics: Silhouette Coefficient Calinski-Harabasz index Bonferroni-Dunn test average rank
Baselines: DK++ k-FED FFCM-avg1 FFCM-avg2 Fed-SC FedSC
Builds on: k-means++ DBI index
Closest papers by meaning
- Enhancing Clustered Federated Learning: Integration of Strategies and Improved Methodologies (ICLR 2025 · similarity 0.88)
- Efficient Asynchronous Federated Learning with Prospective Momentum Aggregation and Fine-Grained Correction (AAAI 2024 · similarity 0.87)
- Clustered Federated Learning via Gradient-based Partitioning (ICML 2024 · similarity 0.87)
- OmniFC: Rethinking Federated Clustering via Lossless and Secure Distance Reconstruction (NeurIPS 2025 · similarity 0.86)
- Interaction-Aware Gaussian Weighting for Clustered Federated Learning (ICML 2025 · similarity 0.86)
- Federated Graph-Level Clustering Network (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.