Graph neural network variants in traffic forecasting: A review

Abstract

The growth of inhabitants and vehicles on one side and limited infrastructure often make traffic congested. Intelligent Traffic System (ITS) is one of the applications in machine learning to manage traffic intelligently. One field in ITS is traffic prediction, where machine learning learns from historical data to forecast future traffic conditions. Natively, the traffic network is in the graph structure where a place is connected to other places through a roadway. The typical machine learning architecture learns from Euclidean data, whereas the characteristic of the graph is non-Euclidean; hence the graph was manipulated to fit on the architecture. Graph Neural Network (GNN) is a recent machine learning architecture designed to learn from graph structure directly. Each node can learn the embedding of the nodes and predict other node information through a message passing mechanism. GNN has some variants, and they are combined with other existing architectures due to the complexity of the problem and the requirement to learn the spatial and temporal features of the traffic network. This article reviews GNN variants and various techniques in traffic forecasting problems.