Source-linked AI summary
Graph Neural Network for Traffic Forecasting: A Survey
Weiwei Jiang, Jiayun Luo
TL;DR
Traffic forecasting is challenging because it involves large, high-dimensional data and graph-based spatiotemporal structure. This paper comprehensively reviews GNN applications for traffic forecasting and reports superior performance over previous approaches on road traffic flow and speed forecasting tasks.
Problem
Traffic forecasting involves large data volumes and high dimensionality, while moving-average models cannot adequately handle graph-based spatiotemporal forecasting.
Method
The paper presents a comprehensive review of GNN applications for traffic forecasting, including rapidly developing research and preprints.
Results
The reviewed approaches show superior performance to previous approaches on road traffic flow and speed forecasting tasks.
Takeaways & Limitations
The survey consolidates research on GNN-based traffic forecasting as a comprehensive review of the application area.
Takeaways & Limitations
Moving-average models are not optimal for traffic forecasting problems with a graph-based form.
Abstract
from arXiv · showhide
Traffic forecasting is important for the success of intelligent transportation systems. Deep learning models, including convolution neural networks and recurrent neural networks, have been extensively applied in traffic forecasting problems to model spatial and temporal dependencies. In recent years, to model the graph structures in transportation systems as well as contextual information, graph neural networks have been introduced and have achieved state-of-the-art performance in a series of traffic forecasting problems. In this survey, we review the rapidly growing body of research using different graph neural networks, e.g. graph convolutional and graph attention networks, in various traffic forecasting problems, e.g. road traffic flow and speed forecasting, passenger flow forecasting in urban rail transit systems, and demand forecasting in ride-hailing platforms. We also present a comprehensive list of open data and source resources for each problem and identify future research directions. To the best of our knowledge, this paper is the first comprehensive survey that explores the application of graph neural networks for traffic forecasting problems. We have also created a public GitHub repository where the latest papers, open data, and source resources will be updated.
1. Introduction
Traffic forecasting must capture large-scale, dynamic spatial and temporal dependencies that traditional time-series and grid-based approaches handle imperfectly. This survey synthesizes GNN-based traffic forecasting research, resources, and future directions across transportation settings.
- Research challenge: Traffic forecasting is challenging because it involves high-dimensional data, spatial dependencies beyond nearby areas, seasonal temporal dependencies, and emergency dynamics.
- Limitations of prior approaches: Traditional ARIMA models cannot handle spatiotemporal forecasting problems, while CNN grid representations are not optimal for graph-based traffic problems.
- GNN motivation: GNNs are suited to traffic forecasting because they capture spatial dependency through non-Euclidean graph structures such as road networks.
- Evidence and scope: GNN-based models have demonstrated superior performance to previous approaches on road traffic flow and speed forecasting tasks.
- Survey design: The survey reviews 212 papers from 2018 to 2020, including preprints, and categorizes traffic problems, graph formulations, and models.
- Resources and outlook: It compiles open datasets, software, and code resources while discussing challenges and future directions for GNN-based traffic forecasting.
2. Related Research Surveys
Existing traffic-prediction surveys classify methods, review architectures and preprocessing, and examine factors affecting accuracy. This survey literature identifies GNNs as the frontier of deep learning-based traffic prediction while addressing their specific application to traffic forecasting.
- The reviewed literature indicates that post-2019 state-of-the-art deep-learning traffic-prediction models are based on GNNs, which remain the state-of-the-art technique.
- Prior surveys categorize traffic-prediction methods into statistics-based, machine-learning, deep-learning, reinforcement-learning, and transfer-learning approaches.
- Earlier surveys report that encoder-decoder LSTM models combined with graph-based methods achieve state-of-the-art traffic prediction performance.
- Related reviews cover model architectures, spatiotemporal feature selection, graph construction, and comparisons with recurrent, temporal-convolutional, sequence-to-sequence, and adversarial networks.
- Sample size and prediction time horizon significantly influence prediction accuracy in deep-learning transport studies.
- No existing survey focuses specifically on applying GNNs to traffic forecasting, motivating a dedicated review of this research area.
3. Problems
The survey categorizes traffic forecasting problems by traffic-state level and reviews how graph neural networks address their complex spatial and temporal dependencies. It covers road, regional, and station-level settings alongside diverse forecasting targets and data representations.
- Problem categorization: Traffic forecasting problems are categorized by road-level, region-level, and station-level traffic states.The survey further breaks problems down by predicted traffic characteristics and groups remaining types under “other problems.”
- Graph representations: Road-network data use road segments or sensors as graph nodes, with connectivity or spatial proximity describing dependencies.GPS trajectories are mapped to the road network, whose topology can contain hundreds or thousands of road segments.
- Graph representations: Station-level problems represent metro or bus stations as graph nodes, with metro lines or bus routes describing spatial dependencies.These station graphs may contain tens or hundreds of stations.
- Graph representations: Regional problems use regular or irregular regions as graph nodes, with inter-region dependencies extracted from land-use purposes.The survey lists regional taxi, bike, and other flow problems among the reviewed applications.
- Forecasting targets: The surveyed applications include traffic flow, speed, travel time, congestion, arrival time, taxi and bike flow, and traffic demand forecasting.The survey lists only problems with GNN-based solutions in its problem table.
- Modeling challenges: Traffic forecasting is difficult because spatial dependencies are complex, nonlinear, and potentially nonlocal, while earlier models have structural limitations.Multivariate time-series models capture only linear relationships, whereas CNNs are bounded to Euclidean structures and cannot model road or subway topology.
- GNN motivation: GNNs address non-Euclidean transportation structures by learning from node and edge attributes and, for dynamic settings, automatically learning graphs from data.The survey also discusses super-graphs and sub-graphs for hierarchical traffic problems.
- Applications: Accurate traffic-flow prediction supports congestion control, traffic-light control, vehicular clouds, and related transportation applications.The survey considers road-level, region-level, and station-level traffic flow problems separately.
4. Graphs and Graph Neural Networks
This section organizes traffic graphs and graph neural networks for forecasting, formalizes graph-based traffic forecasting, and surveys graph construction choices across transportation settings.
- Survey contribution: The survey organizes graph and adjacency-matrix construction approaches and briefly introduces GNN structures frequently used in traffic forecasting.It adds attention-based and FNN-based subtypes of spatiotemporal GNNs and proposes a general design pipeline for future studies.
- Traffic graph formulation: A traffic graph G = (V, E, A) represents nodes, edges, adjacency, and node features χt ∈ R^N×d containing traffic states.N denotes the number of nodes, while d denotes the number of traffic state variables.
- Traffic forecasting formulation: Graph-based traffic forecasting learns y = f(χ; G) from historical traffic states, with an extended formulation y = f(χ, ε; G) incorporating external factors.Single-step forecasting predicts the next time step, whereas multiple-step forecasting predicts traffic states several steps later.
- Graph construction: The survey classifies traffic graphs into road-level, region-level, and station-level graphs according to transportation-domain structure.Road-level examples include sensors, road segments, intersections, and lanes; region-level examples include irregular, regular, and origin-destination graphs.
- Adjacency matrix construction: Adjacency matrices are categorized as road-based, distance-based, similarity-based, or dynamic, with choices controlling how spatial dependencies are represented.Matrices may be predefined static graphs or dynamically learned from continuously evolving data.
5. Open Data and Source Codes
The survey summarizes open data and source-code resources for GNN-based traffic forecasting. These resources support graph formulation, follow-up benchmarking, and replication of prior solutions.
- The section compiles open data suitable for GNN-related traffic forecasting studies.
- The listed data support graph structures used to formulate different forecasting problems.
- The survey lists GNN-related code resources for replicating previous solutions as baselines.
5.1. Open Data
The survey organizes open traffic data into graph-related, historical traffic, and external data, describing how each supports graph construction or forecasting. It also notes practical trade-offs involving coverage, preprocessing, sparsity, and accessibility.
- The survey categorizes data into graph-related, historical traffic, and external data.
- Graph-related data: Transportation network data directly define static graphs, with infrastructure elements as nodes and connections as edges.
- Historical traffic data: Traffic sensor data provide node attributes directly but require imputation or denoising after hardware faults and cover only sensor-equipped locations.
- Historical traffic data: GPS trajectories offer low-cost, wide coverage and can yield flow or speed after map matching, but they contain no direct traffic information.
- Historical traffic data: Trip records can provide speed, demand, and multiple graph-based problems, whereas traffic reports are less used because they are spatially and temporally sparse.
- Open data resources are valuable for benchmarking, but historical data remain harder to source because of privacy, transmission, and storage requirements.
5.2. Open Source Codes
The survey identifies general deep-learning frameworks, GNN-specific libraries, and released traffic-forecasting implementations as open-source resources for research and replication.
- Common Python deep-learning frameworks include TensorFlow, Keras, PyTorch, and MXNet.
- GNN-specific libraries include DGL, pytorch geometric, and Graph Nets.
- Authors have released implementations for traffic flow, speed, demand, and other problems, summarized in Tables 7–10.
- TensorFlow and PyTorch are the two most frequently used frameworks among these open-source projects.
5.3. State-of-the-art Performance
The survey summarizes state-of-the-art traffic-prediction performance using comparable results from frequently used datasets. It cautions that differing datasets, preprocessing, and evaluation settings limit direct comparison and that the results may become outdated.
- Different datasets, subsets, preprocessing methods, and train/validation/test splits make reported results difficult to compare.
- The survey therefore summarizes only comparable results for the most frequently used datasets.
- RMSE, MAE, and MAPE are among the evaluation metrics used for traffic forecasting.
- Lower RMSE or MAE indicates better prediction performance.
- Table 11 summarizes state-of-the-art performance and reports a default prediction period of 60 minutes unless otherwise specified.
- Because research on GNN traffic forecasting continues to grow, the listed results are not guaranteed to be the latest.
6. Challenges and Future Directions
The survey identifies unresolved traffic forecasting challenges and discusses future directions for applying graph neural networks, including opportunities specific to graph-based modeling.
- The survey discusses challenges specific to GNNs alongside broader traffic forecasting problems.
- GNNs improve forecasting performance but do not resolve existing traffic prediction challenges.
- Future directions draw on traffic forecasting research while highlighting special opportunities enabled by GNNs.
6.1. Challenges
The survey identifies challenges involving data, graph structure, multi-task modeling, scalability, changing transportation networks, and model interpretation. These issues constrain practical deployment and remain incompletely solved.
- Data and graph quality: Inaccurate or outdated road-network and graph data can reduce forecasting accuracy.
- Data and graph quality: Missing, sparse, and noisy data compromise forecasting, while most surveyed models rely on processed high-quality datasets.
- Data and graph quality: Congestion can reduce deep neural network performance, but sufficient anomaly data remain difficult to collect for training.
- Multi-task forecasting: Multi-task prediction is difficult because different traffic tasks may require different graph structures, while most models train one task at a time.
- Scalability and deployment: Large-scale GNN computation often requires node and edge subsets, limiting results to selected network portions.
- Dynamic networks and interpretation: Static graphs cannot fully handle changing roads, transit infrastructure, and points of interest, while longer data collection increases practical expense and difficulty.
- Dynamic networks and interpretation: Interpretability remains especially difficult for GNN-based traffic forecasting because heterogeneous traffic data complicate model explanation.
6.2. Future Directions
The survey proposes standardized resources, richer transportation graphs, and advanced learning techniques as directions for addressing data, graph, generalization, and computational challenges. It also reviews early attempts supporting these directions.
- Resources and benchmarking: A centralized repository with standardized formats, graph data, version tracking, public code, ranked results, and long records could support objective model comparison.
- Resources and benchmarking: Standardized benchmarks remain difficult because graph data use varied formats and multiple similarity graphs can be built from the same traffic data.
- Traffic graph design: Transportation knowledge graphs are proposed to extract hidden domain knowledge from multi-source heterogeneous traffic data and leverage traffic semantics.
- Advanced learning techniques: GAN-based data augmentation and AutoML are proposed to address limited data and computational requirements in graph-based forecasting.
- Advanced learning techniques: Transfer learning can move knowledge from data-rich graphs to graphs with less historical data, including unseen highway regions.
- Advanced learning techniques: Meta-learning is proposed for multi-task prediction involving multiple graphs by learning relationships between tasks or data samples.
7. Conclusion
The paper presents a comprehensive survey of GNN applications in traffic forecasting across road-, region-, and station-level problems. It reviews major GNN families, catalogs open resources, and identifies challenges and future research directions.
- The survey covers road-level, region-level, and station-level traffic problems and graphs.
- It discusses recurrent GNNs, convolutional GNNs, and graph autoencoders.
- The paper provides a collection of open datasets and code resources for traffic forecasting.
- It identifies challenges and future directions for follow-up research.