Federated Clustering and Semi-Supervised learning: A new partnership for personalized Human Activity Recognition

Federated Learning (FL) is currently studied by several research groups as a promising paradigm for sensor-based Human Activity Recognition (HAR) to mitigate the privacy and scalability issues of classic centralized approaches. However, in the HAR domain, data is non-independently and identically distributed (non-IID), and personalization is one of the major challenges. Federated Clustering has been recently proposed to mitigate this issue by creating specialized global models for groups of similar users. While this approach significantly improves personalization, it assumes that labeled data are available on each client. In this work, we propose SS-FedCLAR, a novel HAR framework that combines Federated Clustering and Semi-Supervised learning. In SSFedCLAR , each client uses a combination of active learning and label propagation to provide pseudo labels to a large amount of unlabeled data, which is then used to collaboratively train a Federated Clustering model. We evaluated SS-FedCLAR on two well-known public datasets, showing that it outperforms existing semi-supervised FL solutions while reaching recognition rates similar to fully-supervised FL approaches. (c) 2022 Elsevier B.V. All rights reserved.

Automated summary

Federated Learning (FL) is a promising paradigm for sensor-based Human Activity Recognition (HAR) to address privacy and scalability issues. However, non-IID data and personalization pose challenges in the HAR domain. This work proposes SS-FedCLAR, a novel framework combining Federated Clustering and Semi-Supervised learning to enhance recognition rates.