Less Federation, More Personalization: One-Shot Multi-View Clustering for Federated Fall Detection

Abstract

Federated learning is a natural fit for sensor-based fall detection, where data are privacy-sensitive and highly non-IID across subjects. A two-stage framework is proposed in which each client first learns label-free representations using FMMVCC, a multi-view clustering encoder that combines reconstruction, cross-view consistency, contrastive learning, and prototype-based objectives. The resulting encoders are then aggregated via a single round of FedAvg into a shared frozen feature extractor, decoupling representation learning from downstream classification. Experiments are conducted on the FallTL dataset, a wearable sensing benchmark evaluated under both binary (Fall/ADL) and 16-class (activity × Fall/ADL) granularities. Results show that FMMVCC achieves strong representation quality, outperforming four self-supervised baselines in binary classification (84.15% ACC) and providing competitive performance in the fine-grained setting. On top of these representations, the personalized closed-form classifier consistently outperforms FedAvg, FedProx, and shared analytic baselines by 8-13pp, and matches or exceeds FedProto in the 16-class setting while requiring no communication during inference. Finally, cross-client alignment metrics are shown to be weak predictors of downstream performance, while class-separability measures correlate more reliably with accuracy.