Federated ensemble-learning for transport mode detection in vehicular edge network

Abstract

Transport Mode detection has become a crucial part of Intelligent Transportation Systems (ITS) and Traffic Management Systems due to the recent advancements in Artificial Intelligent (AI) and the Internet of Things (IoT). Accurately predicting a person’s mode of transportation was challenging for many years until the computational power of smartphones and smartwatches expanded dramatically over time. This is a result of the numerous sensors built within smart devices, which enable the worldwide cloud server to acquire sensory data and anticipate a person’s method of transport using multiple machine learning models. Currently, all smart devices and vehicular edge devices are interconnected by Vehicular Edge Networks (VEN). However, as the data are shared globally, the security of an individual’s data is questioned, and hence a significant portion of the population is still unwilling to share their sensory data with the global cloud server. Also, the processing time for the massive amount of sensory data should be considered. In this paper, we present a distributed method, Federated Ensemble-Learning in VEN, in which a vast amount of data is used to train the model while the training data is kept decentralized. Federated Ensemble-Learning (FedEL), a hybrid approach, is proposed to enhance the performance of federated strategies. In addition, a majority voting ensembling method has been developed as part of the federated strategy to determine the mode of transportation of local customers. Two machine learning algorithms, XGBoost and Random Forest, and one deep learning technique Multi-Layer Perceptron (MLP) are trained with data from each local client. A prediction is then maintained based on a majority vote among the three models. The class with the most votes is taken into account, while the others are discarded. The FedEL technique has been shown to be highly effective on the TMD dataset, with an accuracy of 94-95% for the 5- second window dataset and 98-99% for the half-second window dataset, based on extensive testing.