Edge-enabled Asynchronous Federated Learning using DAG for Human Mobility Prediction

Accurately predicting human mobility patterns is crucial for a variety of applications including transportation, smart city planning and public health. The existing Machine Learning (ML) models for predicting human mobility are centralized and raises privacy concerns of the individual users data. The emerging Federated learning (FL) paradigm provides privacy-preserving ML solution that enables the training of a global model by coordinating several devices to perform model training without sharing their raw data. However, the requirement of obtaining synchronous model updates from different devices to calculate a global model is still a challenge. Although Directed Acyclic Graph based asynchronous federated learning is explored in the recent works, the methods for model aggregation and scalability issues for overall learning of mobility is challenging as the number of devices participating in the FL process continue to increase. This work propose Edge enabled Directed Acyclic Graph based Asynchronous Federated Learning (EDAG-AFL) framework to manage and integrate model updates across a network of edge servers. This approach eliminates the bottlenecks of centralized systems by enabling asynchronous and scalable learning, where each edge server specializes in local data while benefiting from the collective insights of the entire network. The simulation results demonstrate that the performance EDAG-AFL is comparable to the existing federated learning models while also effectively addressing privacy, synchrony, and scalability challenges.