Federated Topology Preserving Domain Adaptation for Human Activity Recognition

Federated Learning enables decentralized training of machine learning models by collaborating with multiple clients without sharing their data. However, limited labeled data in clients' domains can lead to overfitted local models and compromise the performance of the global model. To overcome this, we propose a novel approach by adopting a single source and multiple targets, considering variations in the style and movements of each client as distinct domains within Human Activity Recognition (HAR). The Modified Topology Preserving Domain Adaptation (modTPDA) is utilized to align the topological structure and preserve the underlying data manifold. Integrating modTPDA with the FL aims to adapt public source data in the server to various target domains with private data, accounting for their data topology. The proposed approach effectively addresses the issue of limited labeled data in Federated Learning for Human Activity Recognition.