A Communication-Efficient Machine Learning Framework for the Internet of Vehicles

The tremendous volume of data produced in the Internet of Vehicles (IoV) presents an opportunity to utilize machine learning (ML) for developing several intelligent services. However, this abundance of data may also give rise to significant communication challenges when attempting to efficiently implement centralized learning (CL). To address this challenge, we propose a novel approach to utilize Mahalanobis distance (MD) metric for sampling the data produced by the IoV. In the proposed approach, vehicles transmit their data to the nearby roadside infrastructure (RSI), which then samples valuable data from the available data using MD metric and relays it to the central server for necessary learning tasks. The sampling ratio for every RSI is determined to satisfy the estimated network delay required to transmit the data to the central server. The proposed method of transmitting only the valuable data, offers the huge potential to establish a communication-efficient learning network for the IoV. Initially, the learning performance of the proposed system is evaluated on synthetic data, followed by testing it on real data to demonstrate its effectiveness. To validate the efficacy of the proposed approach, alternative sampling approaches and several sampling ratios are taken into account and the performance is compared accordingly.