Advancements of Graph Neural Networks in Urban Traffic

Prediction

Jiangnan Liu

and Shipeng Xu

International Engineering College, Xi’an University of Technology, Xi’an, China

College of Agricultural and Biological Sciences, Dali University, Dali, China

Keywords: Graph Neural Networks, Traffic Prediction, Intelligent Transportation Systems.

Abstract: Accurate traffic flow and travel speed prediction are essential for intelligent transportation systems. A

particular kind of deep learning model called Graph Neural Networks (GNNs) is made especially to deal with

graph-structured data, such as road networks. Road networks include intricate relationships and dependencies

that can be precisely captured by them, which makes them an excellent choice for large-scale traffic flow and

journey speed prediction. The use of GNNs in predicting urban traffic is reviewed in this work. We focus on

methods for addressing spatiotemporal dependencies, including Spatiotemporal Graph Neural Networks (S-

GNNs), Temporal Graph Convolutional Networks (T-GCNs), and attention-based techniques. Furthermore,

we discuss Deep GNNs for enhancing traffic prediction accuracy, as well as the application of GNNs

combined with the Internet of Things (IoT) for emergency traffic planning. Despite the substantial potential

of GNNs in traffic prediction, there is a lack of systematic exploration and comprehensive analysis regarding

their applicability in diverse urban environments, the optimization of real-time prediction capabilities, and

their integration with urban planning and management strategies.

1 INTRODUCTION

With the rapid urbanization process, traffic

congestion has emerged as a significant challenge

confronting cities globally. Traffic congestion results

in diminished efficiency of urban operations,

increased fuel consumption, and atmospheric

pollution (Sharma et al., 2023). Accurate traffic flow

and travel speed prediction are paramount for the

design and implementation of Intelligent

Transportation Systems (ITS) (Sharma et al., 2023;

Khorami et al., 2023; Piccialli et al., 2024). ITS can

optimize traffic flow control systems, mitigate road

congestion, enhance road usage efficiency, and foster

environmental sustainability. Traditional machine

learning methods have made significant contributions

to traffic flow prediction, but they are not very good

at capturing complex spatial and temporal

interactions (Piccialli et al., 2024). One kind of deep

learning model created especially to handle graph-

structured data are Graph Neural Networks (GNNs).

They are useful for large-scale traffic flow and road

https://orcid.org/0009-0004-6634-0101

https://orcid.org/0009-0002-0921-5318

segment travel speed prediction because they can

accurately reflect the intricate linkages and

dependencies found in metropolitan road networks

(Khorami et al., 2023; Piccialli et al., 2024).

GNNs offer an effective means to model the

inherent intricate relationships and interactions

within transportation networks. Among these, the

temporal and spatial interdependence issues in traffic

prediction are of significant importance. Luo et al.

have proposed Spatiotemporal Graph Neural

Networks (S-GNNs) as a traffic prediction method

(Luo et al., 2024). S-GNNs investigate the nonlinear

relationships between variables while concurrently

accepting a variety of traffic data inputs. According

to Khorami et al., the temporal and spatial

interconnectedness can be captured by mixing two or

more models (Khorami et al., 2023). One graph

neural network model that combines a Graph

Convolutional Network (GCN) and a Gated

Recurrent Unit (GRU) is called the Temporal Graph

Convolutional Network (T-GCN). In T-GCN,

temporal dependencies are captured by the GRU and

Liu, J. and Xu, S.

Advancements of Graph Neural Networks in Urban Trafﬁc Prediction.

DOI: 10.5220/0012902200004508

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Engineering Management, Information Technology and Intelligence (EMITI 2024), pages 62-66

ISBN: 978-989-758-713-9

spatial aspects are captured by the GCN. Another

example is the integration of GNNs and a Generative

Adversarial Network (GAN) to achieve efficient

traffic management using generative

reasoning(Piccialli et al., 2024). Alourani et al. have

proposed a method using a dual attention neural

network to handle complex situations caused by

outside factors like weather and accidents. The

method uses bidirectional Long Short-Term Memory

(LSTM) units to capture the temporal dependencies

between features and GNNs to represent spatial

features (Alourani et al., 2023). To address the issue

of extracting as much effective information as

possible from nodes, Huang et al. have proposed a

multimodal spatiotemporal convolutional method

that integrates multiple spatiotemporal modules

(Huang et al., 2022).

In addition to addressing the challenges of

temporal and spatial dependencies, GNNs can also be

applied to other facets of traffic prediction. For

example, based on the characteristics of the

interconnections between roads, a deep GNN method

has been proposed to optimize existing GNN models

for traffic prediction (Sharma et al., 2023). For traffic

prediction in scenarios with missing data, Ru et al.

have proposed a method based on GNNs for traffic

prediction using small sample data (Ru et al., 2020).

In terms of optimizing urban emergency

transportation, integrating GNNs with Internet of

Things (IoT) edge computing based on dynamic

graph structures can significantly enhance the

accuracy of traffic prediction (Sun et al., 2022).

Although GNN technology has demonstrated

substantial potential in traffic prediction, there is still

a dearth of systematic exploration and comprehensive

analysis regarding its applicability in diverse urban

environments, the optimization of real-time

prediction capabilities, and its integration with urban

planning and management strategies.

This article is structured in the following manner.

First, we will list the uses of GNNs in intelligent

transportation along with specifics of their

approaches in Section 2. Next, we will examine

GNNs' possible applications in intelligent

transportation in more detail in Section 3 and provide

a clear research framework for upcoming needs. In

conclusion, Section 4 offers a synopsis of the entire

piece.

2 METHOD

2.1 GNNs for Addressing

Spatiotemporal Dependencies

2.1.1 Spatiotemporal Graph Neural

Networks

In the study exploring the application of GNNs to

urban traffic forecasting, the work of Luo et al.

highlights the capability of S-GNNs in handling the

heterogeneity and complexity of traffic data (Luo et

al., 2024). Their study proposes a spatiotemporal

graph neural network model that can accept multiple

traffic data sources as input and deeply explore the

nonlinear correlations among these variables. By

creating a spatiotemporal directed graph, the model is

able to assign distinct attention weights to

neighboring area nodes of the target node in addition

to capturing the sample features at each time step and

aggregating the neighborhood information of nodes

using graph convolution. This approach demonstrates

significant advantages in improving the accuracy of

both short-term and long-term traffic forecasting,

particularly in reducing prediction errors, proving its

effectiveness and feasibility in practical applications.

Further enhancing the performance of traffic flow

forecasting models, the study by Khorami et al.

introduces a Decomposed Temporal Self-Attention

Multi-Layer Graph Convolutional Network (DTSA-

3GCN) (Khorami et al., 2023). This research

effectively addresses the issue of accuracy

degradation in long-term forecasting of traditional T-

GCN models by incorporating Singular Value

Decomposition (SVD), Self-Attention (SA)

mechanism, and Temporal Multi-Layer Graph

Convolutional Networks. The model significantly

improves the performance of GNNs by optimizing the

adjacency matrix and utilizing low-dimensional data.

Khorami et al.'s work not only showcases the

significant improvement in traffic flow forecasting

accuracy by deeply analyzing the spatial and temporal

dependencies in traffic data but also provides new

perspectives and methods for urban traffic

management and planning (Khorami et al., 2023).

2.1.2 Attention Mechanisms-Based

Techniques

For traffic forecasting models, the ability to consider

external influencing factors, such as weather and

incidents, is crucial. In the literature, Alourani et al.

successfully demonstrate the significant role of dual

attention mechanisms in enhancing the accuracy of

Advancements of Graph Neural Networks in Urban Trafﬁc Prediction

traffic flow forecasting through their research

(Alourani et al., 2023). Their study combines GNNs

and Bidirectional Long Short-Term Memory

(BiLSTM) networks, introducing a novel dual

attention mechanism that focuses on capturing both

the spatial characteristics and temporal dependencies

of traffic data simultaneously. Specifically, the spatial

attention module concentrates on analyzing the traffic

flow relationships between different geographical

locations, while the temporal attention module

investigates the evolving traffic patterns over time.

Through this approach, Alourani et al.'s model

successfully improves the traffic flow prediction

accuracy of specific road segments while considering

the impact of weather factors, confirming the

importance of dual attention mechanisms for

understanding and predicting complex traffic systems

(Alourani et al., 2023).

2.1.3 Multimodal Data Fusion

Techniques

In recent years, multimodal data fusion techniques

have shown significant promise in improving traffic

forecasting models' performance. Huang et al.

propose a novel traffic flow forecasting framework

based on multimodal spatiotemporal convolution

(Huang et al., 2022). By fusing data from

heterogeneous sources (e.g., vehicle detector data,

social media data), the framework deeply analyzes

the relationships among different data sources using

an adaptive spatiotemporal convolution module,

mixed jump-diffusion Ordinary Differential Equation

(ODE) spatiotemporal convolution module, and

multimodal spatiotemporal fusion module, leading to

more accurate predictions of traffic flow and behavior

patterns. This not only provides new perspectives for

urban traffic management and planning but also

showcases the potential of deep learning approaches

in handling complex spatiotemporal dependencies.

2.2 Deep GNNs for Enhanced Traffic

Forecasting Accuracy

Faced with complex road network structures and

diverse traffic conditions, simple GNN frameworks

fall short in making accurate traffic predictions,

leading to the proposal of Deep GNN approaches for

enhanced accuracy in complex traffic forecasting.

The STGGAN model was proposed by Sharma et al.

for real-time traffic speed prediction (Sharma et al.,

2023). The spatiotemporal properties of traffic speed

data are extracted using the STGNN model using

multi-level spatiotemporal graph analysis. The deep

learning model, spatial feature aggregation, and

feature extraction are the three main elements that

make up the model. There are numerous noteworthy

aspects in the STGNN model. First of all, it captures

intricate interactions between geographical and

temporal aspects by integrating graph structures and

spatiotemporal data. Second, it uses a Gated

Recurrent Unit (GRU) layer to efficiently capture

temporal correlations in traffic data. Finally, a Graph

Attention Network (GAT) layer is incorporated into

the model to represent spatial dependencies and make

use of previous road network knowledge.

Furthermore, incorporating attention mechanisms

into GNNs, also a type of Deep GNN approach, can

enhance traffic forecasting accuracy. Based on GPS

data from taxis, Ru et al. present a model for

anticipating traffic operational conditions in urban

road networks by employing a limited number of

crucial segments (Ru et al., 2020). To determine the

crucial sections and attain high prediction accuracy,

the model makes use of a graph neural network with

an attention mechanism. The suggested approach

lowers the price of gathering traffic data and is

economical. The suggested model examines how

each section contributes to the prediction of vehicle

speed using the attention mechanism. The attention

coefficients are calculated using the softmax function.

The model also incorporates a graph attention

network to consider the correlation between segments

in the road network.

2.3 GNNs Combined with IoT for

Emergency Traffic Planning

The Internet of Things (IoT) plays a pivotal role in

urban traffic management, with a vast number of IoT

devices deployed across cities, capable of collecting

massive amounts of traffic data. The combination of

GNNs and IoT has a wide range of applications,

particularly in IoT system intrusion detection (Zhou

et al., 2022). Sun et al. design a dynamic graph

structure that works in conjunction with a GNN

algorithm, enabling rapid traffic forecasting using

small and local data collected in real-time from IoT

devices in urban areas, thus addressing the issue of

emergency traffic planning in cities (Sun et al., 2022).

This dynamic graph structure is capable of expanding

inside a local mobile communication network, taking

into account both temporal and geographical

information. It uses a graph representation of the road

network, with nodes standing in for intersections and

edges for roads. Each node and edge are timestamped,

indicating the time of traffic flow data collection. To

capture dynamic changes in traffic flow, the approach

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

employs time series modeling techniques. It

represents historical traffic flow data of each node as

a time series and uses GNNs to learn patterns in the

time series. This method's primary benefit is its

compatibility with Internet of Things devices,

utilizing small and local datasets for real-time

forecasting. This is particularly useful in small-scale

cities while also being scalable to larger datasets and

more complex traffic networks.

3 DISCUSSIONS

Although GNNs excel at capturing the complex

temporal and spatial relationships in traffic networks,

leading to improved prediction accuracy and reduced

data usage costs, particularly in large and medium-

sized cities, the application of GNNs in traffic

prediction is not without its challenges. Firstly, GNN

models are often black-box models, making it

challenging to interpret the reasons behind their

predictions. This can be a problem for safety-critical

applications, where it is important to understand how

the model makes its decisions. Secondly, the

performance of GNN models can be limited and

affected by imbalanced data in urban traffic, such as

the occurrence of localized peak traffic flow on urban

roads. Thirdly, GNN models are sensitive to changes

in the graph structure, which can lead to performance

degradation when the traffic network changes. There

are also several other factors that need to be

considered when applying GNNs to traffic prediction

and intelligent traffic management. In addition to the

urban Internet of Things as the basic infrastructure of

smart cities, to maximize the synergistic effect, it is

imperative to investigate the most effective ways to

integrate GNN models with other technologies.

Despite the challenges and limitations discussed

above, GNNs hold great promise for revolutionizing

traffic prediction and intelligent traffic management.

Future research should focus on developing

interpretable GNN models, exploring data-efficient

GNN models, and investigating the integration of

GNNs with other technologies. Interpretable

GNNs can help to address the black-box nature of

GNNs by providing explanations for their predictions.

The predictions of GNN models can be explained by

interpretability techniques like Shapley Additive

Explanations (SHAP) or Local Interpretable Model-

Agnostic Explanations (LIME). A unique machine

learning strategy called domain adaptation seeks to

enhance the model's performance on the target

domain by adjusting the difference between training

and testing data from various distributions (Farahani

et al., 2021). Therefore, combining GNN with domain

adaptation will help address the limitation of GNN

models due to the imbalanced distribution of traffic

data. Integrating GNNs with other technologies such

as attention mechanism (Qiu et al., 2022) can help to

further improve the performance and robustness of

traffic prediction and intelligent traffic management

systems. For example, GNNs can be integrated with

reinforcement learning to develop adaptive traffic

signal control systems, or with generative adversarial

networks to generate realistic traffic scenarios for

training and testing. In addition, graph structure

adaptive GNN models will also be an important

research direction in the future. Graph structure

adaptive GNN models enable them to handle changes

in the graph structure. This can be achieved by using

dynamic graph convolution or graph attention

mechanisms, which can capture changes in the graph

structure and adjust the model weights accordingly.

The optimization directions mentioned above for

GNNs will enable them to play a more crucial role in

future urban traffic prediction.

4 CONCLUSIONS

This paper provides a comprehensive review of recent

research on the application of GNNs in urban traffic

prediction, demonstrating the unique advantages and

potential of GNN technology. Our primary

contributions lie in exploring and analyzing different

GNN models, such as S-GNNs, Deep GNNs, and

GNNs integrated with the IoT, revealing the potential

and effectiveness of these models in improving the

accuracy of urban traffic prediction and optimizing

urban emergency traffic systems. GNNs are capable

of effectively capturing the complex dependencies in

traffic data, enhancing the accuracy of traffic flow

and vehicle speed predictions. Particularly, when

considering the spatial characteristics of road

networks and the temporal variations of traffic flow,

GNNs demonstrate their unique advantages.

However, despite the significant potential shown by

GNNs in urban traffic prediction, our research also

highlights their limitations in interpretability,

applicability, and other aspects. We believe that

future research should focus on optimizing the real-

time prediction capabilities of GNN models,

improving the interpretability of models, and

exploring their application in a wider range of urban

environments. Additionally, integrating the latest

Internet of Things technologies to further enhance the

effectiveness of GNNs in urban emergency traffic

planning is also an important research trend.

Advancements of Graph Neural Networks in Urban Trafﬁc Prediction

AUTHORS CONTRIBUTION

All the authors contributed equally, and their names

were listed in alphabetical order.

REFERENCES

Alourani, A., Ashfaq, F., Jhanjhi, N. Z., & Khan, N. A.

2023. BiLSTM- and GNN-Based Spatiotemporal

Traffic Flow Forecasting with Correlated Weather Data.

Journal of Advanced Transportation, 2023, 1–17.

Farahani, A., Voghoei, S., Rasheed, K., & Arabnia, H. R.

2021. A Brief Review of Domain Adaptation. In R.

Stahlbock, G. M. Weiss, M. Abou-Nasr, C.-Y. Yang, H.

R. Arabnia, & L. Deligiannidis (Eds.), Advances in

Data Science and Information Engineering (pp. 877–

894). Springer International Publishing.

Huang, X., Lan, Y., Ye, Y., Wang, J., & Jiang, Y. 2022.

Traffic Flow Prediction Based on Multi-Mode Spatial-

Temporal Convolution of Mixed Hop Diffuse ODE.

Electronics, 11(19), 3012.

Khorami, R., Kortetjärvi, F., & University, H. 2023. Graph

Neural Network for Traffic Flow Forecasting: Does an

enriched adjacency matrix with low dimensional data

enhance the performance of GNN for traffic flow

forecasting?

Luo, M., Dou, H., & Zheng, N. 2024. Spatiotemporal

Prediction of Urban Traffics Based on Deep GNN.

CMC, 78(1), 265–282.

Piccialli, F., Canzaniello, M., Chiaro, D., Izzo, S., & Qi, P.

2024. GRAPHITE — Generative Reasoning and

Analysis for Predictive Handling in Traffic Efficiency.

Information Fusion, 106, 102265.

Qiu, Y., Wang, J., Jin, Z., Chen, H., Zhang, M., & Guo, L.

2022. Pose-guided matching based on deep learning for

assessing quality of action on rehabilitation

training. Biomedical Signal Processing and Control, 72,

103323.

Ru, X., Xu, X., Zhou, Y., & Yang, C. 2020. Critical

Segments Identification for Link Travel Speed

Prediction in Urban Road Network. Journal of

Advanced Transportation, 2020, 1–10.

Sharma, A., et al. 2023. A Graph Neural Network (GNN)-

Based Approach for Real-Time Estimation of Traffic

Speed in Sustainable Smart Cities. Sustainability,

15(15), 11893. https://doi.org/10.3390/su151511893

Sun, B., Geng, R., Shen, T., Xu, Y., & Bi, S. 2022. Dynamic

Emergency Transit Forecasting with IoT Sequential

Data. Mobile Networks and Applications.

Zhou, X., Liang, W., Li, W., Yan, K., Shimizu, S., & Wang,

K. I. -K. 2022. Hierarchical Adversarial Attacks

Against Graph-Neural-Network-Based IoT Network

Intrusion Detection System. IEEE Internet of Things

Journal, 9(12), 9310–9319.

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence