Network Federated Learning for Real Time Traffic Predictions in Internet of Vehicles Scenarios

The project aims to present a theoretical and algorithmic framework for Network Federated Learning (NFL) in decentralized collections of local datasets, where the datasets are structurally related through some specific similarity measure. The similarity can be based on functional relationships, statistical dependencies, or spatiotemporal proximity. In our case, we considered the mobility model of vehicles to model the movement of vehicles within an area to better analyse the dynamics of traffic within the city. In order to formulate NFL, our methodology uses a Generalized Total Variation (GTV) minimization which is the extension of existing federated multi-task learning methods. The approach we presented is adaptable and works with different models. Moreover, for local models that result in convex problems, we provide precise conditions on the local models and their network structure such that the algorithm learns nearly optimal local models. The analysis reveals an interesting interplay between the convex geometry of local models and the (cluster-) geometry of their network structure. Means, the shape of the local models which is determined by their convex geometry, and the structure of the network determined by their cluster geometry, can influence the quality of the learned local models. We then compare the results of our algorithm with centralized federated learning (FL) model. The analysis shows our algorithm for NFL has quite better performance with respect to centralized FL.