GCN-GRU for Junction-Level Traffic Flow Prediction: A Systematic Review and Comparative Synthesis with CNN and LSTM/GRU

Urban traffic congestion is one of the most pressing challenges of the twenty-first century, impacting quality of life, economic productivity, mobility, and environmental sustainability.[1,2] Accurate traffic flow forecasting has therefore become central to Intelligent Transportation Systems (ITS), particularly at junctions where multiple flows converge and diverge, creating bottlenecks and congestion.[3,4] Junction-level prediction is especially critical since these points concentrate the complexity of road networks and often determine overall traffic dynamics.[5] Traditional models such as regression-based methods, Kalman filters, and ARIMA struggled with the nonlinear and high-dimensional nature of modern traffic data.[6] The learning method offered advances through CNNs for spatial feature extraction and RNN variants (LSTM, GRU) for temporal dynamics.[5,7] However, key challenges remain: traffic networks are inherently non-Euclidean, with road segments and intersections forming irregular graphs rather than grids; sensor data is often sparse or missing; and models trained in one domain may not generalize well under shifting traffic patterns. These limitations highlight the need for approaches that jointly capture temporal sequences and network topology while remaining robust to incomplete and heterogeneous data. This gap is filled by the development of Graph Convolutional Networks (GCNs),[8] which allow representation learning on graph-structured data directly. The resulting GCN-GRU hybrid models can concurrently capture temporal dynamics of traffic flow and spatial dependencies across irregular road networks when combined with recurrent architectures like GRU. GCN-GRU is positioned as one of the most promising methods for junction-level traffic prediction by this synthesis. To map the research landscape, synthesize recent advancements, and critically compare GCN-GRU approaches against CNN and LSTM/GRU models, the study performs a Systematic Literature Review (SLR) in conjunction with bibliometric analysis. This review offers a comprehensive view of the current state of research and potential future directions in junction-level traffic prediction by combining quantitative bibliometric mapping with qualitative synthesis.[9]

The development of traffic prediction methods demonstrates a distinct paradigm shift from statistical to deep learning. For short-term prediction, classical models like ARIMA and Kalman filtering provided interpretable answers, but they were unable to keep up with the growing complexity of real-world, multi-source traffic data. Deep learning models have been adopted more quickly as a result of the growth of large-scale traffic sensing infrastructure, which includes GPS, loop detectors, and IoT-enabled vehicle sensors. Although CNN-based models excel at capturing spatial dependencies, they are unable to accurately depict irregular urban road networks due to their reliance on grid-structured inputs. In contrast, LSTM and GRU models are very good at predicting temporal sequences and identifying long-term dependencies in traffic flow, but they are unable to explicitly model the spatial relationships between intersections. By directly learning spatial representations from graph-structured road networks, GCN-based techniques overcome this gap. The GCN-GRU hybrid architecture is specifically designed for junction-level prediction tasks by combining gated recurrent units for temporal sequence modeling with graph convolution for spatial correlation learning.[10]

To illustrate this methodological evolution, the review summarizes the comparative strengths and weaknesses of CNN, LSTM/GRU, and GCN-GRU models in traffic flow prediction. This tabular synthesis demonstrates how each model family addresses certain challenges while leaving others unresolved, thereby highlighting the rationale for GCN-GRU hybrid approaches.

This systematic review article that analyzes existing studies on GCN-GRU-based traffic flow prediction and compares them with CNN and LSTM/GRU models. To ensure methodological transparency and reproducibility, this review adheres to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) framework. Using keyword combinations like "traffic flow prediction," "junction-level traffic," "GCN," "GRU," "CNN," and "LSTM," extensive searches were carried out in Web of Science, Scopus, and IEEE Xplore. A total of 432 studies published between 2017 and 2025 were found in the first search. 87 papers were kept for the final synthesis after duplicates were eliminated and relevance was checked using keywords, abstracts, and titles.[38]

In parallel, bibliometric mapping was carried out to examine citation clusters, co-authorship patterns, and keyword co-occurrence using VOSviewer. A thorough grasp of the research landscape is ensured by this dual approach, which combines quantitative breadth and qualitative depth through systematic literature synthesis backed by bibliometric analysis.

Additionally, Lens.org's dataset-level statistics, which included 3,846 scholarly works with over 97,000 citations, 315 works cited by patents, and 579 citing patents, validated the scope and significance of this field of study. This highlights the field's industrial and applied significance in addition to its academic maturity.[39]

The PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) framework is followed in this review, guaranteeing methodological rigor, reproducibility, and transparency. Web of Science, Scopus, and IEEE Xplore were the databases that were searched.

Keywords like "traffic flow prediction," "junction-level traffic," "GCN," "GRU," "CNN," and "LSTM" were combined in the search strategy.

Bibliometric analysis tool: VOSviewer.

Dimensions analyzed:

Dataset-level statistics (Lens.org):

Significance: Confirms both academic maturity and practical relevance of junction-level traffic prediction research.[41]

This study may be affected by selection bias, as the datasets used for evaluation may not fully represent all traffic conditions or geographic regions. Dataset coverage is another potential limitation, since the chosen datasets may not include variations across different times, seasons, or unusual traffic events. These factors could limit the generalizability of the proposed model to other traffic environments. Future work should include more diverse datasets to address these threats and strengthen the validity of the findings.[42]

Recent advances in deep learning have significantly transformed junction-level traffic flow prediction, evolving from traditional time-series models to sophisticated spatiotemporal architectures. In this context, GCN-GRU hybrid models have emerged as a powerful alternative to standalone CNN or LSTM/GRU approaches by jointly capturing spatial dependencies and temporal dynamics. Bibliometric analysis reveals evolving trends in authorship, prevalent keywords, and citation patterns, reflecting the growing interest in enhanced and attention-based models. Despite these advances, research gaps remain in model generalization, explainability, and real-time deployment, pointing to promising future directions for further exploration.[43,44]

The co-authorship analysis shows that Chinese and U.S. institutions have made the most contributions to the field of GCN-based traffic prediction. There are many strong national research clusters, but there aren't many international collaborations. This means that there needs to be more cross-continental engagement and data sharing. This kind of collaboration would make it easier to use models in different cities.

While the co-occurrence network highlights thematic concentrations and methodological trends, the underlying co-authorship patterns reveal relatively weak international linkages. This limited cross-regional collaboration may restrict access to diverse traffic datasets, which are essential for validating models across different urban contexts. Without stronger dataset sharing and global cooperation, the transferability of traffic prediction models across cities remains constrained, potentially reducing their effectiveness in heterogeneous real-world environments.

The keyword co-occurrence mapping shows three main research groups:

Cluster 1: GCN, graph neural networks, and modeling of space and time.

Group 2: LSTM, GRU, and predicting time series.

Cluster 3: CNN, modeling based on images, and a grid-based representation.

This shows that there has been a big change from CNN and RNN models to graph-based architectures. The increasing popularity of GCN-GRU keywords is a clear sign that this model is becoming the best way to predict junction levels. Fig. 2 shows the keyword co-occurrence overlay visualization, highlighting how research emphasis has shifted from broader themes such as "deep learning," "application," and "survey" toward more recent and specialized topics like "graph convolutional network (GCN)," "dependency," and "spatial correlation" between 2021 and 2023. This progression reflects the field's movement from general methodological foundations to advanced modeling of non-Euclidean traffic structures. However, the fragmented distribution of keywords across clusters also indicates limited cross-regional integration, reinforcing the finding from co-authorship analysis that weak international collaboration may hinder dataset sharing and reduce the transferability of models across diverse urban contexts.

Citation analysis identifies seminal GCN-based models like DCRNN and ST-GCN as central nodes, which are heavily cited in later research. Recent studies that combine GCN with GRU show an increase in citation momentum, which shows that more people are recognizing their predictive power. In contrast, models that only use CNN and LSTM show fewer citations, which means they are becoming less important in cutting-edge traffic prediction research.

Fig. 3 presents the keyword density visualization, where brighter areas highlight the most frequently cited and co-occurring terms such as "graph convolutional network (GCN)," "graph," "deep learning," and "application." These dense regions indicate the dominant methodological focus of recent research, with strong emphasis on graph-based modeling and spatiotemporal dependencies. However, the density remains uneven across clusters, reflecting the influence of a limited number of research hubs. Combined with the weak international linkages observed in co-authorship networks, this suggests barriers to dataset sharing and knowledge transfer across regions, potentially constraining the generalizability of models to cities with different traffic infrastructures and mobility patterns. Collaboration patterns in traffic prediction research, such as combining graph-based spatial learning with recurrent temporal models (e.g., CNN-GRU-LSTM hybrids or GCN-based approaches), enhance model adaptability to diverse datasets. Models trained using multi-source data or collaborative frameworks generally exhibit better transferability across different traffic environments. This suggests that integrating diverse collaboration patterns improves robustness when applying models to unseen traffic scenarios.

When it comes to junction-level traffic prediction tasks, the comparative evidence clearly favours GCN-GRU models. Although CNNs are good at modeling spatial features, their ability to do so is constrained by the presumption of Euclidean grid structures. Road network topology is not explicitly taken into account by LSTM and GRU, despite their superiority at capturing sequential dependencies.

By combining the advantages of both paradigms, GCN-GRU hybrids achieve improved accuracy and reduced RMSE on benchmark datasets like METR-LA and PEMS-BAY. Importantly, because graph-based representations enable information to spread among connected nodes, GCN-GRU exhibits resilience against missing data and sensor failure. GCN-GRU is a strong option for scalable, practical ITS deployment because of these benefits.