Federated Learning for Human Activity Recognition with Environmental and Subject-Level Awareness

Abstract

In human activity recognition studies, the location of the sensors and environmental conditions are determined with precision during the data collection process, and this positioning can directly affect the model performance. In addition, the individual characteristics of the participants while performing the activities stand out as a separate problem element that makes the generalizability of the system difficult. In order to overcome these two problems, the federated learning method was applied in the presented study to obtain a model that adapts to different environmental conditions and has a high generalization capacity among users. Within the scope of this study, studies were carried out on two different scenarios. In these scenarios, client types were determined as environmentbased and subject-based and the studies were carried out in this direction. In addition, the effect of aggregation functions on system performance was examined in these two scenarios. The experimental results show that FedAdam achieves superior performance in environment-based systems, while FedYogi achieves superior performance in subject-based systems.