This is the authors' abstract. We don't add key points for this paper.
Federated learning has shown its great applicability in intelligent transportation systems, where prediction models can be trained across regions or cities without leaking raw data. However, current federated learning approaches often ignore energy consumption, while energy consumption is playing a pivotal role in sustainability of transportation systems. In this paper, we propose a Two-Phase Client Selection strategy for federated learning (FedTPCS) in traffic flow prediction, aiming to minimize the total energy consumption of clients for participation while considering device and data heterogeneity. First, to tackle device heterogeneity, we leverage K-means clustering to group clients based on their computing power and geographic distance. We strategically select the clustered group with the lowest average cost that is the combination of energy consumption and latency. Second, to tackle data heterogeneity, we leverage affinity propagation clustering based on cosine similarity of model update vectors to divide the selected clients into several subgroups of similar clients. We evaluate the performance of the proposed FedTPCS algorithm on two public datasets. Compared to FedAvg, FedAEB and Greedy algorithms, the FedTPCS algorithm reduces cost by up to 56%, 30%, and 20% under the PeMS dataset, and 50%, 28%, and 18% under the Highways England dataset, respectively.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: