Federated-Learning-for-Human-Mobility-Models
Thanks to the proliferation of smart devices, such as smartphones and wearables, which are equipped with computation, communication and sensing capabilities, a plethora of new location-based services and applications are available for the users at any time and everywhere. Understanding human mobility has gain importance to offer better services able to provide valuable products to the user whenever it's required. The ability to predict when and where individuals will go next allows enabling smart recommendation systems or a better organization of resources such as public transport vehicles or taxis. Network providers can predict future activities of individuals and groups to optimize network handovers, while transport systems can provide more vehicles or lines where required, reducing waiting time and discomfort to their clients. The representation of the movements of individuals or groups of mobile entities are called human mobility models. Such models replicate real human mobility characteristics, enabling to simulate movements of different individuals and infer their future whereabouts. The development of these models requires to collect in a centralized location, as a server, the information related to the users' locations. Such data represents sensitive information, and the collection of those threatens the privacy of the users involved. The recent introduction of federated learning, a privacy-preserving approach to build machine and deep learning models, represents a promising technique to solve the privacy issue. Federated learning allows mobile devices to contribute with their private data to the model creation without sharing them with a centralized server. In this thesis, we investigate the application of the federated learning paradigm to the field of human mobility modelling. Using three different mobility datasets, we first designed and developed a robust human mobility model by investigating different classes of neural networks and the influence of demographic data over models' performance. Second, we applied federated learning to create a human mobility model based on deep learning which does not require the collection of users' mobility traces, achieving promising results on two different datasets. Users' data remains so distributed over the big number of devices which have generated them, while the model is shared and trained among the server and the devices. Furthermore, the developed federated model has been the subject of different analyses including: the effects of sparse availability of the clients; The communication costs required by federated settings; The application of transfer-learning techniques and model refinement through federated learning and, lastly, the influence of differential privacy on the modelβs prediction performance, also called utility
