Advanced Traffic Flow Prediction Using Graph

Neural Networks
Dushyant Sharma Jeet Singh Sahil
Department of Computer Science Department of Computer Science Department of Computer Science
& Engineering & Engineering & Engineering
Chandigarh University Chandigarh University Chandigarh University
Punjab, India Punjab, India Punjab, India
dushyant.e16465@[Link] jeetsogarwal10@[Link] jangrasahil2021@[Link]

Subham Chandel Shiven Chauhan

Department of Computer Science Department of Computer Science
& Engineering & Engineering
Chandigarh University Chandigarh University
Punjab, India Punjab, India
subhamchandel403@[Link] shivu.9r@[Link]

Abstract—Traffic congestion remains a critical issue in urban face challenges related to land availability, urban planning
areas, contributing to substantial economic losses and environ- restrictions, and long implementation timelines.
mental degradation. Accurate prediction of traffic flow is essential With advancements in sensor technology, large-scale traffic
to improve transportation efficiency and alleviate congestion.
Traditional statistical models such as ARIMA have laid the data collection has become more accessible, paving the way
foundation, but recent advancements in machine learning have for predictive modeling in traffic management. Traffic flow
significantly improved predictive capabilities. Among these, deep prediction has emerged as a crucial research area, enabling
learning techniques, particularly graph neural networks (GNNs), data-driven decision-making that offers a cost-effective and
have gained prominence due to their ability to model intricate scalable alternative to physical infrastructure expansion. Short-
spatial and temporal relationships within road networks.
This paper presents a comprehensive survey of deep learning
term traffic forecasting, typically spanning 5 to 60 minutes into
approaches for traffic flow prediction, with an emphasis on GNN- the future, is particularly valuable for optimizing transportation
based models. We systematically review and categorize existing networks. However, unlike conventional time-series forecast-
studies, compare various deep learning architectures, and identify ing, traffic prediction is highly dependent on spatial corre-
emerging trends, persistent challenges, and promising future lations and external influences such as weather conditions,
directions. By offering a detailed examination of the strengths
and limitations of GNNs in traffic forecasting, this review
special events, and holidays. These factors introduce additional
aims to guide future research and development in intelligent complexity, making accurate forecasting a challenging task.
transportation systems. Among various predictive modeling techniques, deep neural
Index Terms—component, formatting, style, styling, insert networks (DNNs) are increasingly being explored due to their
ability to capture complex, nonlinear traffic patterns. The most
I. I NTRODUCTION widely used DNN architectures in traffic prediction include
Convolutional Neural Networks (CNNs), Recurrent Neural
Traffic congestion has become a serious issue in major Networks (RNNs), and Feedforward Neural Networks (FNNs).
metropolitan areas, leading to significant economic losses and Despite the increasing number of studies leveraging deep
environmental degradation. In Australia alone, the estimated learning for traffic forecasting, the diversity of hybrid model
financial impact of traffic congestion rose from $12.8 billion in structures has made it challenging to establish a unified un-
2010 to approximately $16.5 billion in 2015, with projections derstanding of this field. Furthermore, comprehensive surveys
reaching $30 billion by 2030. Traditional congestion mitiga- specifically analyzing deep neural network-based traffic pre-
tion strategies, such as limiting the number of vehicles on the diction methods remain scarce, leaving a gap in the literature
road—an approach used in Singapore—may not be feasible in for systematic evaluation.
countries with underdeveloped public transportation systems. This paper aims to bridge that gap by presenting a structured
Expanding road networks is another potential solution, but it and up-to-date review of deep learning models for traffic
comes with extremely high costs. For instance, constructing forecasting, with a particular emphasis on Graph Neural Net-
a single lane of road in New Jersey, USA, can cost around works (GNNs). Our target audience includes researchers and
$220,490 per mile. In addition, such physical expansions often practitioners interested in applying advanced neural network
architectures to traffic prediction. The paper is organized as
follows:
• Section 2 provides a historical perspective on traffic
prediction research and outlines key developments in
modeling techniques, from early statistical methods to
modern deep learning approaches.
• Section 3 explores the three primary deep learning
models used in traffic forecasting: Convolutional Neural
Networks (CNNs), Recurrent Neural Networks (RNNs),
and Feedforward Neural Networks (FNNs). It discusses
their unique strengths, limitations, and suitability for
capturing spatiotemporal traffic dynamics.
• Section 4 serves as the core of our study, categorizing
and analyzing 37 state-of-the-art research contributions. Fig. 1.
This section is further divided into two parts: Section 4.1,
which examines the datasets used in traffic forecasting, Zhu et al. conducted another relevant survey, focusing on
including their scale, features, and availability; and Sec- big data analytics in traffic prediction rather than the predic-
tion 4.2, which provides an in-depth analysis of the deep tive models themselves. While their work provides valuable
learning models applied, focusing on their architectural insights into data handling and management, it does not offer
innovations and performance comparisons. a comprehensive analysis of deep learning models or how they
• Section 5 addresses the challenges in traffic prediction re-
are applied to real-world traffic systems. Our review takes a
search. Section 5.1 highlights current issues such as data more balanced approach by covering both datasets and neural
sparsity, model generalization, and real-time deployment network architectures, offering a more holistic understanding
constraints, while Section 5.2 explores future research of the field.
directions, including the integration of external factors,
One of the key aspects we emphasize in this study is the
multi-modal data fusion, and the role of explainable AI.
comparison of different hybrid deep neural network imple-
• Section 6 concludes the paper with a summary of key in-
mentations. With the widespread availability of deep learning
sights, implications for intelligent transportation systems,
libraries such as Keras, PyTorch, and TensorFlow, researchers
and potential avenues for advancing deep learning-based
have increasingly developed hybrid models that integrate mul-
traffic forecasting research.
tiple architectures to capture different aspects of traffic data.
A. Comparison with Existing Surveys For example, hybrid models often combine CNNs for spatial
One of the most notable literature surveys in this domain feature extraction with RNNs for temporal sequence modeling,
was conducted by Vlahogianni et al., focusing primarily on and in some cases, even GNNs for capturing topological rela-
the challenges of traffic prediction rather than specific model tionships. Despite the prevalence of such approaches, there has
architectures. However, their work only covered research from been limited systematic analysis comparing their effectiveness
2006 to 2013, missing the widespread adoption of deep neural and adaptability across diverse traffic scenarios. To the best
networks. Additionally, their classification of models, based of our knowledge, this review is the first attempt to provide a
on broad categories such as statistical models and neural structured comparison of various hybrid deep learning models
networks, is now outdated given the dominance of deep in traffic prediction.
learning in modern traffic forecasting. Their study also lacked Through this review, we aim to provide a clear understand-
insights into recent developments like hybrid architectures, ing of the current state of traffic prediction research, highlight
graph-based methods, and the growing availability of large- existing gaps, and offer insights into future developments in
scale traffic datasets that have transformed the field in recent the field, particularly focusing on the role of Graph Neural
years. Networks in improving traffic forecasting accuracy and scala-
A more recent review by Nagy and Simon provides an bility.
overview of traffic prediction models but lacks depth in model II. BACKGROUND
comparison. Their taxonomy is based on a limited set of
criteria, such as whether a model integrates environmental A. Problem Formulation
factors or captures spatial properties. Our work, in contrast, Traffic prediction involves developing a learnable function
provides a more detailed comparison framework, analyzing that utilizes historical traffic data to forecast future traffic
both models and datasets across eleven distinct evaluation conditions. The two primary types of traffic data used in
points. Additionally, unlike their survey, our study includes prediction models are traffic flow and traffic speed. Traffic
a discussion on future challenges in traffic prediction, offering flow refers to the total number of vehicles detected at a
a more forward-looking perspective that aligns with current given location over a specific time period, while traffic speed
technological advancements and research priorities. represents the average velocity of vehicles at that location
within the same period. In this paper, the term “traffic” is
used interchangeably to refer to both traffic flow and speed.
Mathematically, traffic prediction can be defined as:
ŷt+T0 = f (Xt−T1 , Xt−T , ..., Xt )
where:
• yt represents the observed traffic at time t.
• ŷt denotes the predicted traffic at time t.
• T refers to the input sequence length, i.e., how many past
time steps are used for prediction.
• T0 is the prediction horizon, indicating how far into the
future the prediction is made.
• f is a learnable function responsible for generating pre- Fig. 2.
dictions based on historical data.
• L is the loss function, which evaluates the accuracy of
the predictions. 2) Machine Learning Models: To overcome the shortcom-
• θ represents the optimal set of parameters for function f ,
ings of statistical models, researchers began adopting machine
found by minimizing the prediction error. learning (ML) techniques, which can automatically learn
patterns from data rather than relying on manually predefined
The choice of function f , loss function L, and optimal
parameters. Unlike classical models, ML approaches dynam-
parameters θ depends on the model used for traffic prediction.
ically adjust their parameters during training, making them
In the following sections, we explore different modeling
more adaptable to complex traffic conditions.
approaches and discuss the evolution of traffic prediction
Among various ML techniques, neural networks became
techniques over time.
the most widely used due to their ability to extract meaningful
B. Evolution of Traffic Prediction Models features automatically. Traditional ML models such as k-
Nearest Neighbors (k-NN) and Support Vector Regression
The field of traffic prediction has evolved over the past
(SVR) were also applied to traffic prediction but lacked the
five decades, with models categorized into three major groups:
flexibility and scalability of neural networks. One of the
classical statistical models, machine learning models, and
deep learning models.
1) Classical Statistical Models: One of the earliest and
most widely used statistical methods for traffic prediction is
the Autoregressive Integrated Moving Average (ARIMA)
model. It was first applied to traffic forecasting by Ahmed
and Cook, followed by Levin and Tsao, who discovered
that ARIMA(0,1,1) provided statistically significant results in
freeway traffic prediction.
Several researchers further improved ARIMA-based mod-
els: Fig. 3.
• Subset ARIMA was introduced by Lee and Fambro,
demonstrating increased stability and accuracy. earliest applications of neural networks for traffic forecast-
• ARIMAX, an extended version incorporating external ing was introduced by Dougherty et al. Over time, various
variables, was applied by Williams to account for the enhancements were proposed:
influence of upstream traffic on downstream areas. • Genetic algorithms were used by Vlahogianni et al. to
• Seasonal ARIMA was utilized by Williams and Hoel to optimize neural network parameters.
analyze traffic patterns in the U.S. and U.K. • Bayesian probability-based models were introduced by
• Space-Time ARIMA models were explored by Kamar- Zheng and Lee, combining multiple neural networks to
ianakis and Prastacos to capture spatiotemporal correla- improve prediction accuracy.
tions in traffic data. • Time-delay neural networks were explored by Zhong
Despite their widespread use, classical statistical models have et al. to incorporate temporal dependencies in traffic
notable limitations. They assume traffic data is stationary forecasting.
and linear, which makes them ineffective in handling com- • Hybrid models combining neural networks with statisti-
plex, nonlinear traffic patterns. Additionally, these models cal smoothing techniques were proposed by Chan et al.
were designed for smaller datasets, which is no longer ideal to improve training efficiency.
given the rapid growth of real-time traffic data collected from Although machine learning models outperformed statistical
sensors and IoT devices. approaches, early implementations were shallow and had lim-
ited computational power, restricting their effectiveness in period marks significant advancements in deep learning ap-
capturing complex traffic dynamics. However, advancements plications for traffic forecasting. Additionally, seminal works
in hardware, coupled with improved theoretical frameworks, on traditional statistical and machine learning models were
paved the way for deep learning in traffic prediction. included to provide historical context.
3) Emergence of Deep Learning: By the mid-2010s, deep To improve the robustness of the search process, Boolean
learning models began to dominate the field of traffic pre- operators (AND, OR) and citation chaining were used to
diction. Deep Neural Networks (DNNs) consist of multiple discover additional relevant articles referenced within selected
layers of interconnected neurons, allowing them to capture papers. Duplicate studies and non-English publications were
both spatial and temporal dependencies in traffic data more excluded. Emphasis was placed on peer-reviewed articles,
effectively. conference proceedings, and journal publications to ensure
Some of the most commonly used deep learning models quality and relevance.
include: Furthermore, studies that lacked clear experimental design,
• Recurrent Neural Networks (RNNs) for modeling se- evaluation metrics, or dataset transparency were excluded from
quential dependencies in time-series traffic data. the final selection. Papers focusing exclusively on non-urban
• Convolutional Neural Networks (CNNs) for extracting environments, vehicle detection, or unrelated transportation
spatial features from traffic maps and sensor networks. aspects were also filtered out. This curated selection strategy
• Feedforward Neural Networks (FNNs) for learning ensured a focused and in-depth analysis of traffic forecasting
general patterns in traffic behavior. techniques based on deep learning.
• Hybrid models combining multiple deep learning archi- The final corpus consisted of 37 research articles, each
tectures to leverage their respective strengths. subjected to a detailed evaluation framework comprising ar-
Deep learning models have demonstrated remarkable ac- chitecture design, dataset usage, input features, spatiotempo-
curacy in traffic prediction by capturing intricate patterns that ral modeling techniques, evaluation metrics, and comparative
traditional methods could not. However, they also come with performance.
challenges:
• High Data Requirements: Deep models require vast
amounts of training data that adequately represent all
possible traffic conditions.
• Computational Complexity: Training deep neural net-
works is computationally intensive and time-consuming,
especially when using hybrid models.
• Interpretability Issues: Due to their complex architec-
tures, deep learning models act as ”black boxes,” making
it difficult to interpret their decision-making processes.
Fig. 4.
III. M ETHODOLOGY
To conduct a comprehensive review of deep learning models
for traffic flow prediction, a structured methodology was B. Inclusion and Exclusion Criteria
employed to ensure the selection, analysis, and comparison A rigorous and methodical selection process was carefully
of relevant research studies. This section outlines the search applied to ensure both the relevance and academic quality of
strategy, inclusion and exclusion criteria, classification frame- the reviewed literature. The following inclusion and exclusion
work, and analytical approach utilized in this study. criteria were employed:
A. Literature Search Strategy Inclusion Criteria:
The research articles included in this review were sourced • Research articles focusing on deep learning models for

from reputable academic databases, including IEEE Xplore, traffic prediction.

Springer, ScienceDirect, and Google Scholar. The literature • Studies utilizing real-world traffic datasets for model

search was conducted using a combination of keywords such evaluation.

as: • Peer-reviewed journal articles and conference proceed-

• “Traffic flow prediction using deep learning”

ings.
• Papers that present comparative analyses of different
• “Graph Neural Networks for traffic forecasting”
• “Machine learning in intelligent transportation systems”
predictive models.
• “Spatiotemporal modeling for traffic data” Exclusion Criteria:
• “Hybrid deep learning models for traffic forecasting” • Studies focusing solely on traditional statistical methods
• “Short-term traffic prediction using neural networks” without deep learning integration.
To ensure the review remains current and relevant, the primary • Non-peer-reviewed papers, preprints, and articles lacking
focus was on studies published between 2014 and 2024, as this technical depth.
 • Research relying exclusively on synthetic datasets with- A. Evaluation Metrics
out real-world applicability.
To ensure consistent and fair performance comparisons, the
reviewed studies employed several standard metrics:
C. Classification Framework
• Mean Absolute Error (MAE) – Measures the average
To facilitate a structured analysis, the selected studies were magnitude of prediction errors.
categorized based on three primary dimensions: • Root Mean Square Error (RMSE) – Penalizes larger
1) Model Type: The deep learning architecture utilized, in- errors more significantly than MAE.
cluding Convolutional Neural Networks (CNNs), Recur- • Mean Absolute Percentage Error (MAPE) – Shows
rent Neural Networks (RNNs), Graph Neural Networks prediction accuracy as a percentage, offering easy inter-
(GNNs), and hybrid models. pretability.
2) Dataset: The traffic datasets employed in the studies, • R-squared (R2 ) – Represents the proportion of variance
such as PeMS (California freeway data), METR-LA, and explained by the model.
TaxiBJ.
3) Performance Metrics: The evaluation criteria used, B. Model Comparison
including Mean Absolute Error (MAE), Root Mean
Square Error (RMSE), and R2 score, to assess model The reviewed literature includes diverse deep learning ar-
accuracy and effectiveness. chitectures:
• CNN-based Models: Effective for spatial feature extrac-
tion from traffic maps and sensor layouts.
• RNN and LSTM-based Models: Strong in capturing
temporal dependencies and sequence-based patterns.
• Hybrid Architectures (e.g., CNN+LSTM): Combine
spatial and temporal learning for more accurate forecast-
Fig. 5. Possible inconsistency scenario in traditional Active Replication. ing.
• Graph Neural Networks (GNNs): Model traffic net-
works as graphs, providing superior performance in com-
plex, irregular road topologies.
D. Comparative Analysis Approach
Following the classification, a comprehensive comparative C. Dataset Usage
analysis was conducted to examine:
• The strengths and limitations of different deep learning The studies analyzed span across several widely-used
models. benchmark datasets:
• The influence of dataset characteristics on model perfor- • METR-LA – Traffic speed data from Los Angeles.
mance. • PEMS-BAY – California-based traffic flow dataset.
• The effectiveness of hybrid architectures in capturing • PEMSD4 and PEMSD8 – From the Caltrans Perfor-
spatial and temporal dependencies in traffic data. mance Measurement System.
• Custom Urban Datasets – Specific to cities in Asia and
E. Identifying Research Gaps Europe, focusing on real-time flow.
Real-world datasets, particularly those with detailed spatiotem-
Through systematic literature analysis, key challenges and
poral structure, consistently supported more accurate and
gaps were identified, including the need for enhanced model
generalizable models.
interpretability, incorporation of external influencing factors
(e.g., weather conditions, special events), and optimization of
computational efficiency. These insights informed the discus- D. Performance Comparison Table
sion on future research directions in traffic prediction.

TABLE I
IV. E VALUATION AND C OMPARATIVE A NALYSIS P ERFORMANCE C OMPARISON OF D EEP L EARNING M ODELS ON
METR-LA DATASET
To provide a comprehensive evaluation of deep learning
models used for traffic flow prediction, this section presents Model MAE ↓ RMSE ↓ MAPE (%) ↓ R2 ↑
CNN 3.45 7.01 12.3 0.82
a detailed comparative analysis based on model performance, LSTM 3.20 6.72 11.5 0.85
architectural variations, dataset diversity, and evaluation met- CNN-LSTM 2.98 6.34 10.8 0.88
rics. The goal is to identify the most effective model types GNN (ST-GCN) 2.67 5.98 9.7 0.90
under varying data and environmental conditions.
 methodologies effectively. A crucial step in this direction is
the development of comprehensive benchmark datasets that
incorporate a wide range of influencing factors. In addition to
spatiotemporal traffic data, integrating external variables such
as weather conditions, accident reports, and real-time social
media feeds will enhance predictive accuracy and robustness.
Furthermore, fostering transparency in this research domain is
essential. Open-source implementations and publicly accessi-
ble model architectures will facilitate reproducibility, acceler-
ate progress, and encourage collaboration among researchers
and practitioners. Additionally, ongoing systematic reviews
and experimental evaluations will provide valuable insights
into emerging trends and innovations in the field.
The continued advancements in traffic flow prediction hold
significant potential for the widespread deployment of real-
time traffic forecasting systems. Such implementations can
play a pivotal role in optimizing traffic management, reducing
congestion, and improving overall transportation efficiency.
By addressing the existing challenges and embracing future
opportunities, this field can drive meaningful progress toward
Fig. 6. Sample visualization of predicted vs. actual traffic flow smarter, more adaptive urban mobility solutions worldwide.
R EFERENCES
E. Visual Evaluation (Graph) [1] Australian Government Bureau of Infrastructure, Transport and Regional
Economics, (2015), “Information Sheet 74 - Traffic and Congestion Cost
F. Insights and Observations Trends for Australian Capital Cities,” viewed Mar. 18, 2020.
The comparative analysis shows that hybrid and graph- [2] BBC News, “Singapore to freeze car numbers,” 2017. [Online]. Avail-
able: [Link] Accessed: Nov. 20,
based models significantly outperform traditional CNN and 2018.
RNN models. Specifically, GNN-based architectures consis- [3] J. Carnegie and A. M. Voorhees, “The Cost of Roadway Construction,
tently achieve lower error margins and better generalization Operations and Maintenance in New Jersey: Phase 1 Final Report,” pp.
557–566, 2016.
on unseen data, primarily due to their ability to model non- [4] H. Yu, Z. Wu, S. Wang, Y. Wang, and X. Ma, “Spatiotemporal Re-
Euclidean spatial relationships. Moreover, models trained on current Convolutional Networks for Traffic Prediction in Transportation
diverse and rich datasets such as PEMS-BAY tend to be more Networks,” Sensors, vol. 17, no. 7, pp. 1–16, 2017.
[5] W. Huang, G. Song, H. Hong, and K. Xie, “Deep Architecture for Traffic
robust to noise and variability. Flow Prediction: Deep Belief Networks with Multitask Learning,” IEEE
In conclusion, while conventional deep learning models Transactions on Intelligent Transportation Systems, vol. 15, no. 5, pp.
continue to be effective, the evolution toward graph-based 2191–2201, Oct. 2014.
[6] Y. Wu and H. Tan, “Short-term Traffic Flow Forecasting with Spatial-
architectures marks a substantial leap in traffic forecasting Temporal Correlation in a Hybrid Deep Learning Framework,” 2016.
accuracy and scalability. These findings reinforce the relevance [Online]. Available: [Link]
of exploring GNNs as a primary tool for future intelligent [7] X. Cheng, R. Zhang, J. Zhou, and W. Xu, “DeepTransport: Learning
Spatial-Temporal Dependency for Traffic Condition Forecasting,” in
transportation systems. Proc. Int. Joint Conf. Neural Netw., 2017, vol. 1709.09585, pp. 1–8.
[8] X. Dai, R. Fu, Y. Lin, L. Li, and F. Wang, “DeepTrend: A Deep Hi-
V. C ONCLUSION erarchical Neural Network for Traffic Flow Prediction,” 2017. [Online].
Traffic flow prediction represents a cost-effective and effi- Available: [Link]
[9] E. I. Vlahogianni, M. G. Karlaftis, and J. C. Golias, “Short-term Traffic
cient approach to mitigating congestion in urban transportation Forecasting: Where We Are and Where We’re Going,” Transportation
networks. This review has provided a comprehensive analysis Research Part C: Emerging Technologies, vol. 43, pp. 3–19, 2014.
of the evolution of traffic prediction methodologies, tracing [10] A. M. Nagy and V. Simon, “Survey on Traffic Prediction in Smart
Cities,” Pervasive Mobile Computing, vol. 50, pp. 148–163, 2018.
the progression from classical statistical models to machine [11] L. Zhu, F. R. Yu, Y. Wang, B. Ning, and T. Tang, “Big Data Analytics
learning techniques and, ultimately, to deep neural networks. in Intelligent Transportation Systems: A Survey,” IEEE Transactions on
We have examined the most commonly employed deep learn- Intelligent Transportation Systems, vol. 20, no. 1, pp. 383–398, Jan.
2019.
ing architectures, outlined their functionality, and explored
how these models extract meaningful patterns from traffic
data. Furthermore, a comparative assessment of state-of-the-
art deep learning approaches for traffic flow prediction has
been presented, along with an in-depth discussion of existing
challenges and potential future research directions.
Moving forward, the advancement of traffic flow prediction
requires a more standardized framework to evaluate novel