Federated Learning for Mobility Applications

Abstract

The increasing concern for privacy and the use of machine learning on personal data has led researchers to introduce new approaches to machine learning. Federated learning is one such a novel privacy-preserving machine learning approach that “brings code to data,” unlike traditional machine learning approaches that “bring data to code.” In addition to improving privacy, federated learning is beneficial for latency-sensitive mobility applications by providing local models. To the best of our knowledge, this article is the first ever to survey mobility-related federated learning solutions, such as traffic-flow prediction, next-location prediction, and point-of-interest recommendation. Our categorization is based on three main questions: Why use federated learning? to identify the motivation to use federated learning; What problems are being addressed? to examine problems that surface with federated learning and how they are solved; and How is federated learning implemented? to account for the solutions implemented by the authors surveyed The selected papers are peer reviewed and published in journals and conferences; they all adopt federated learning as their core approach. We introduce our conceptual model to characterize federated learning solutions and to compare them. In our conceptual model, we define three abstract roles: data generator, learner, and aggregator. We also explain how the work in the selected papers fits into our conceptual model.