Low-overhead Clustered Federated Learning for Personalized Stress Monitoring

Abstract

<p>Advances in Artificial Intelligence (AI) and Wearable Internet of Things (WIoT) are enabling remote health monitoring in everyday settings for early detection and prevention of chronic health problems. Such solutions can be used to augment a conventional physician-centered healthcare system. Stress, as one of the critical health problems, affects individuals adversely in terms of both physical and mental health. Prior studies on stress evaluation utilize a centralized cloud-based approach that combines data from each client for modeling. However, such a centralized approach raises data privacy concerns. To preserve privacy, decentralized federated learning has been proposed as a potential alternative framework. Nevertheless, existing federated learning algorithms have to deal with data heterogeneity; data skewness in each participant can significantly degrade the overall model performance. To tackle this challenge, we present a personalized, low-overhead clustered federated learning algorithm for stress-level recognition. The proposed algorithm outperforms two state-of-the-art baseline algorithms by providing over 7% and 12% increase in accuracy, respectively. We also present a novel cold-start algorithm for new clients who join the trained system. Our results suggest that this cold-start algorithm is robust in terms of individual classification accuracy and total training time.</p>

Federated Clustered Multi-Domain Learning for Health Monitoring

“…domains obtained from the estimation algorithm from our prior work 41 . The multi-domain learning module g(θ ; φ ) : X d i → Y i maps the data labeled with domain tags X d i to the set of labels Y i , where θ is the domain-agnostic parameter and φ is the domain-specific parameter.…”

“…Therefore, we apply K-Means to the latent space data. Since we have no prior knowledge of the number of domains K, we adapt an iterative search algorithm from our prior work 41 : We search for the optimum number of clusters (domains) incrementally. Given a specific domain number NumDomain, we evaluate the clustering quality by computing the average Silhouette score score avg across all N clients.…”

“…Following our prior work 41 , we utilized Stochastic Gradient Descent (SGD) as the training optimizer with a weight decay of 10 −4 and a momentum of 0.9. The maximum communication round is 50, with an early stopping criterion of stopping training after having five consecutive epochs without improved accuracy.…”

“…The maximum communication round is 50, with an early stopping criterion of stopping training after having five consecutive epochs without improved accuracy. We implemented the federated clustering module using the model from our prior work 41 . We trained the ClusterGAN at each client with a batch size of 32 and a learning rate of 5 × 10 −4 for five epochs.…”

“…Cluster-based FL aims at reducing inter-client data heterogeneity by grouping similar patients such that clients within a cluster share similar data distributions and those across different clusters have distinct data distributions. Following our prior work 41 , we performed cluster number estimation, trained federated ClusterGAN, and assigned a cluster id to each client based on the latent embedding from the ClusterGAN. We then applied FedAvg for clients in the same cluster.…”