Definition of a novel federated learning approach to reduce communication costs

Background and Objective: Contemporary Machine Learning approaches (e.g., Deep Learning) need huge volumes of data to build accurate and robust statistical models. Nowadays, very often, such data are collected by distinct and geographically distributed entities and successively transmitted to and stored by centralized nodes that implement the learning process. This practice, however, exposes data to security and privacy risks that may be even unacceptable in those environments regulated by the General Data Protection Regulation (GDPR). Methods: This paper defines a novel Federated Learning approach that avoids the transmission of sensitive data over the network and improves over the classic federated learning schemes by reducing the communication costs. This approach has been validated concerning a healthcare case study that aimed at building a Time-Series based predictive model to identify the level of risk for patients suffering from hypertension. Results: Experimental validation has shown that the proposed approach achieves excellent results both in terms of classification accuracy, superior to the state-of-the-art models with an improvement ranging from 3.01% to 11.09%, and in terms of communication costs with a reduction of about 34%. Conclusion: The analysis of the proposed approach shows promising results in terms of performance and communication cost.