Federated Learning Application in Human Activity Recognition: The Impact of Aggregation Functions

In recently, Human Activity Recognition studies occupy an important place among artificial intelligence problems. Although many successful results have been achieved in numerous studies conducted in this field, data privacy is often overlooked. In this presented study, a Federated Learning application, which prioritizes data privacy, has been used to solve the Human Activity Recognition problem. In this context, a study has been conducted in which four different strategies were tested and the Transformers architecture was used as a model. When the obtained results were examined, it was observed that the Federated Loss-Based Coefficients method, which produced successful results especially in heterogeneous data sets, achieved an average of 96.08% success, obtaining better results compared to other aggregation functions.