Error accumulation and the vanishing gradient problem make temporal neural networks less effective at handling long time series. Consequently, GNN-based models often struggle to produce reliable results for multi-step predictions.
The ability of GNNs to manage spatial correlations is dependent on road topological connections, making them unable to identify geo-parcel-based traffic patterns that are independent of road networks.
GNNs can only capture spatio-temporal dependencies between pairs of nodes, which cannot handle high-order correlations among time series. To address this limitation, multi-graph and hypergraph approaches have been proposed.
The message passing in GNNs is immediate, whilst the spatial message interactions among neighboring nodes are delayed in reality. Indeed, the change of traffic flow at one node will take several minutes (i.e., time delay) to influence its connected neighbors. To address this issue, differential equation models have been used in the literature . It includes both delay effects and continuity into a unified delay differential equation framework. In 2025, Zheng et al. (2025) propose a Delay-aware Directed Graph Attention (DDGA) to delay information propagation between nodes.
Traffic conditions undergo continuous changes. The prediction frequency for traffic flow forecasting may vary based on specific scenario requirements. Thus, adaptive mechanisms are required.
STGNNs neglect the incorporation of the cyclical/periodical patterns that appear in the traffic historical data. For example, the cyclical/periodical patterns of traffic on the same day or hour of each week or season of each year can help improve the accuracy of future traffic predictions. To leverage periodicity information with recognizing that periodic patterns come from prior ones-a phenomenon, Zhang et al. (2024) defined the transitivity of periodicity and designed a Dynamic Graph Convolutional Network grounded in Transmissibility-Periodicity (TPDGCN).
STGNNs provide forecasts without estimates of data and model uncertainty, which are critical for understanding inherent variations of the data and forecast limitations due to a lack of training data. In 2024, Mallick et al. (2024) developed a scalable deep ensemble approach to quantify data and model uncertainties for STGNNs.
The generalization of STGNNs in extended temporal scenarios and cross-city applications remains largely unexplored. In 2024, Wang et al. (2024) evaluated several models and demonstrated performance degradation in existing STGNNs over time due to their limited inductive capabilities. To address this limitation, Wang et al. (2024) proposed a Principal Component Analysis (PCA) embedding approach that allows the models to adapt to new scenarios without retraining.
Graph indistinguishability requires investigations, as graphs learned by existing methods tend to converge to implicit and indistinguishable representations, deviating from the genuine distribution. This problem is due to the lack of three factors within graphs: the intrinsic graph features, the temporal-distinct features, and the node-distinct features. In 2024, Chen et al. (2024) developed a Temporal-Aware Structure-Semantic Coupled Graph Network (TASSGN) that allows for tackling these three factors.
The prior STGNNs ignore the natural hierarchical structure of the traffic road networks and thus fail to capture the macro-spatial dependence of region networks. To address this limitation, Xie et al. (2024) developed a hierarchical GNN for accurate mobile traffic forecasting that fully exploits the cross-regional feature impacts.
STGNNs require prohibitive quadratic computational complexity to capture long-range spatio-temporal dependencies, thus limiting their use to long historical sequences on large-scale road networks in the real world (Han et al., 2024).
The presence of noise in graph nodes can significantly degrade the performance of GNNs (Ding et al., 2024), particularly when local spatial or spectral features are unreliable. To address this issue, several approaches have been proposed. For instance, MRGAT (Ding et al., 2023) introduces multi-scale receptive fields and edge-aware attention mechanisms to enhance local-global feature extraction in noisy environments. AF2GNN (Ding et al., 2022) improves robustness by adaptively combining multiple graph filters and aggregators. MARP (Zhang et al., 2023) employs an adaptive path aggregation mechanism to dynamically learn receptive fields and mitigate the impact of noisy nodes on classification.
Most of existing graph based methods l intricate spatio-temporal dependencies. Neverthless, t not every node within this spatio-temporal graph contributes equally to the modeling of such dependencies. Consequently, the ability to discern and effectively leverage the significant nodes within the graph holds the key to enhancing the accuracy of traffic prediction models.
Fair Use Policy
As this website gives many information that come from my research, please cite my following paper:
B. Remmouche, D. Boukraa, A. Zakharova, T. Bouwmans, M. Taffar, "Long-Term Spatio-Temporal Graph Attention Network for Traffic Forecasting", Expert Systems with Applications, 2025.