Graph neural network in traffic speed prediction: a study on efficiency of training

Abstract

Real-time traffic speed prediction and dynamic traffic control are needed to reduce traffic congestion. In recent years, many examples using one of the deep learning methods “Graph Neural Network (GNN)” have been reported. However, the problem is the long training time and large memory usage when the data size is large. Therefore, the important issue is to train models efficiently while aiming for high accuracy. In this paper, we attempt to reduce the training time using the open data, METR-LA dataset and road traffic data in England, in traffic speed prediction. A sensitivity analysis of the training times and accuracies was conducted when the initial values of the adjacency matrix are manipulatively changed, and it was found that the optimal initial values differ depending on the data. For data in England, the method proposed in this paper reduced training time without sacrificing accuracy compared to previous methods.