Efficiency-Improved Federated Learning Approaches for Time of Arrival Estimation

Abstract-Driven by artificial intelligence and edge computing, federated learning methods have become increasingly advanced and applied in various fields. Specifically, federated learning has been applied in various scenarios in intelligent transportation system (ITS) in which there are privacy issues. Travel time prediction is an important research problem in ITS. In this study, we apply wide-deep-recurrent models in a federated learning framework to predict estimated time of arrival (ETA) of vehicles, which could potentially improve the drivers and passengers’ experience in travel service. We also apply periodic layers updating and representational consistent enhancing strategies in synchronous and asynchronous federated learning to improve the model efficiency. In addition, we propose a synchronous learning rate decay method to enhance the efficiency of federated learning. We integrate federated learning frameworks with wide-deep-recurrent learning and use a ridesharing order records data with multi-dimensional information to predict travel time in a decentralized way to protect personal data privacy. Comparing with the selected baselines, FedAvgPR and FedAsyncPR models can reduce communication cost by about 14.22 % and 44.41 % to reach a target level of accuracy measured by mean absolute percentage error (MAPE). Moreover, synchronous learning-rate decay method not only can reduce the communication cost, but also improve the optimal prediction accuracy by about 1.80 % in synchronous federated learning models. Overall, the optimal accuracies of the proposed models range from 8.42 % to 10.21 %. The proposed federated learning models could be helpful for improving drivers travel experience while tackling the potential data privacy concerns.