Traffic Flow Prediction Based on Federated Learning with Joint PCA Compression and Bayesian Optimization

Traffic flow prediction (TFP) is of great significance in the field of traffic congestion mitigation on the Internet of Vehicle(Iov). To be capable of a trade-off between data privacy protection and accurate prediction, we introduce a training paradigm based on Federated Learning (FL). However, the implementation of federal learning in practice is confronted with high communication and data heterogeneity. In this paper, Principal component analysis (PCA) is introduced to minimize the scale of data transmission on both the client and server. Due to the errors arising from the compression and reversion of the transmission model, we add an additional error term in the local objective function. To address the imbalanced data distribution and to accelerate the federal learning convergence, we then propose a mechanism that incorporates Bayesian optimization to dynamically determine the weights of clients during aggregation. With extensive experiments on real data, it can be demonstrated that communication costs can be minimized by 60-70% while ensuring fewer errors.