JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 1 Graph Neural Network for Traffic Forecasting: A Survey Weiwei Jiang, Jiayun Luo Abstract—Traffic forecasting is important for the success of with the external factors which affect traffic states, e.g. weather intelligent transportation systems. Deep learning models, includ- and holidays. ing convolution neural networks and recurrent neural networks, Both short-term and long-term traffic forecasting problems have been extensively applied in traffic forecasting problems to model spatial and temporal dependencies. In recent years, to for various transport modes are considered in the literature. model the graph structures in transportation systems as well This survey focuses on the data-driven approach, which in- as contextual information, graph neural networks have been volves forecasting based on historical data. The traffic forecast- introduced and have achieved state-of-the-art performance in a ing problem is more challenging than other time series fore- series of traffic forecasting problems. In this survey, we review casting problems because it involves large data volumes with the rapidly growing body of research using different graph neural networks, e.g. graph convolutional and graph attention high dimensionality, as well as multiple dynamics including networks, in various traffic forecasting problems, e.g. road traffic emergency situations, e.g. traffic accidents. The traffic state in flow and speed forecasting, passenger flow forecasting in urban a specific location has both spatial dependency, which may not rail transit systems, and demand forecasting in ride-hailing be affected only by nearby areas, and temporal dependency, platforms. We also present a comprehensive list of open data and which may be seasonal. Traditional linear time series models, source resources for each problem and identify future research directions. To the best of our knowledge, this paper is the first e.g. auto-regressive and integrated moving average (ARIMA) comprehensive survey that explores the application of graph models, cannot handle such spatiotemporal forecasting prob- neural networks for traffic forecasting problems. We have also lems. Machine learning (ML) and deep learning techniques created a public GitHub repository where the latest papers, open have been introduced in this area to improve forecasting data, and source resources will be updated. accuracy, for example, by modeling the whole city as a Index Terms—Traffic Forecasting, Graph Neural Networks, grid and applying a convolutional neural network (CNN) as Graph Convolution Network, Graph Attention Network, Deep demonstrated by [1]. However, the CNN-based approach is not Learning. optimal for traffic foresting problems that have a graph-based form, e.g. road networks. In recent years, graph neural networks (GNNs) have become I. INTRODUCTION the frontier of deep learning research, showing state-of-the- Ransportation systems are among the most important art performance in various applications [2]. GNNs are ideally T infrastructure in modern cities, supporting the daily com- suited to traffic forecasting problems because of their ability muting and traveling of millions of people. With rapid ur- to capture spatial dependency, which is represented using non- banization and population growth, transportation systems have Euclidean graph structures. For example, a road network is become more complex. Modern transportation systems encom- naturally a graph, with road intersections as the nodes and road pass road vehicles, rail transport, and various shared travel connections as the edges. With graphs as the input, several modes that have emerged in recent years, including online GNN-based models have demonstrated superior performance ride-hailing, bike-sharing, and e-scooter sharing. Expanding to previous approaches on tasks including road traffic flow and arXiv:2101.11174v2 [cs.LG] 15 Feb 2021 cities face many transportation-related problems, including air speed forecasting problems. These include, for example, the pollution and traffic congestion. Early intervention based on diffusion convolutional recurrent neural network (DCRNN) [3] traffic forecasting is seen as the key to improving the efficiency and Graph WaveNet [4] models. The GNN-based approach of a transportation system and to alleviate transportation- has also been extended to other transportation modes, utilizing related problems. In the development and operation of smart various graph formulations and models. cities and intelligent transportation systems (ITSs), traffic To the best of the authors’ knowledge, this paper presents states are detected by sensors (e.g. loop detectors) installed the first comprehensive literature survey of GNN-related ap- on roads, subway and bus system transaction records, traffic proaches to traffic forecasting problems. While several relevant surveillance videos, and even smartphone GPS data collected traffic forecasting surveys exist [5], [6], [7], [8], [9], [10], [11], in a crowd-sourced fashion. Traffic forecasting is typically [12], [13], [14], [15], [16], [17], [18], [19], none are GNN- based on consideration of historical traffic state data, together focused. For this survey, we reviewed 212 papers published in the years 2018 to 2020. Additionally, because this is a very Manuscript received ; revised . rapidly developing research field, we also included preprints W. Jiang is with Department of Electronic Engineering, Tsinghua Univer- that have not yet gone through the traditional peer review sity, Beijing, 100084, China. (e-mail: [email protected]). J. Luo is with Department of Statistics, University of California-Los process (e.g., arXiv papers) to present the latest progress. Angeles, Los Angeles, 90024, USA. (e-mail: [email protected]). Based on these studies, we identify the most frequently JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 2 considered problems, graph formulations, and models. We also In [5], three spatiotemporal sequence forecasting (STSF) investigate and summarize publicly available useful resources, categories are highlighted, namely, moving point cloud tra- including datasets, software, and open-sourced code, for GNN- jectory forecasting, STSF on regular grids, and STSF on based traffic forecasting research and application. Lastly, we irregular grids. Both classical and deep learning solutions for identify the challenges and future directions of applying GNNs these problems are reviewed. The scope is beyond the traffic to the traffic forecasting problem. Our aim is to provide a domain and the solution is not limited to GNNs or deep neural comprehensive summary of related work and a useful starting networks. point for new researchers in this field. In addition to this paper, A taxonomy of existing traffic prediction methods, including we have created an open GitHub repository on this topic 1, both traditional and deep learning methods, is presented in [7]. where relevant content will be updated continuously. A series of experiments are conducted using publicly available Our contributions are summarized as follows: datasets for model performance comparison; and the state-of- 1) Comprehensive Review: We present the most compre- the-art methods for prediction tasks including flow, speed, de- hensive review of graph-based solutions for traffic forecasting mand, occupancy, and travel time are summarized. Of GNNs, problems in the past three years (2018-2020). only graph convolutional networks (GCNs) are considered. 2) Resource Collection: We provide the latest comprehen- A broader concept of human mobility is reviewed in [8]. sive list of open datasets and code resources for replication This includes tasks such as next-location prediction, crowd and comparison in future work. flow prediction, and trajectory generation. The authors provide 3) Future Directions: We discuss several challenges and a detailed explanation of different model structures including potential future directions for researchers in this field. the CNN, recurrent neural network (RNN), generative adver- The remainder of this paper is organized as follows. In sarial network (GAN), and attention mechanism. Additionally, Section II, we compare our work with other relevant research rich mobility data sources, including mobile phone data, GPS surveys. In Section III, we categorize the traffic forecasting traces, social media data, and auxiliary data, are identified and problems that are involved with GNN-based models. In Sec- discussed. tion IV, we summarize the graphs and GNNs used in the ML models for traffic prediction are categorized in [10]. reviewed studies. In Section V, we outline the open resources. These include the regression model, example-based models Finally, in Section VI, we point out challenges and future (e.g., k-nearest neighbors), kernel-based models (e.g. support directions. vector machine and radial basis function), neural network models, and hybrid models. In this paper, the GNN (a type II. RELATED RESEARCH SURVEYS of neural network) is not explicitly and separately listed. Both In this section, we introduce the most recent relevant relevant real-world datasets and microscopic traffic simulators research surveys (most of which were published in 2020). are summarized. Useful information about data preprocessing The differences between our study and these existing surveys and model design is presented, as well as the open challenges are pointed out when appropriate. We start with the surveys to inspire future work. addressing wider ITS topics, followed