Source-linked AI summary
Decoupled Dynamic Spatial-Temporal Graph Neural Network for Traffic Forecasting
Zezhi Shao, Zhao Zhang, Wei Wei, Fei Wang, Yongjun Xu, Xin Cao, Christian S. Jensen
TL;DR
Traffic forecasting is challenging because prior STGNNs treat traffic as diffusion while neglecting inherent signals. The paper proposes DSTF and D2STGNN to decouple these components and learn dynamic dependencies, with experiments on four real-world datasets showing consistent improvements over baselines.
Problem
Prior traffic-forecasting methods largely treat traffic data as diffusion signals, neglecting inherent signals and often using incomplete representations of dynamic spatial dependency.
Method
DSTF separates diffusion and inherent signals through residual decomposition and an estimation gate; D2STGNN separately models them and learns dynamic spatial dependencies.
Results
Experiments on four real-world, large-scale datasets show that D2STGNN consistently and significantly outperforms all baselines.
Takeaways & Limitations
Decoupling traffic signals enables subsequent diffusion and inherent models to focus on their respective signal types.
Takeaways & Limitations
The superiority of D2STGNN alone does not isolate the effectiveness of DSTF because performance also depends on temporal and spatial dependency modeling and dynamic topology.
Abstract
from arXiv · showhide
We all depend on mobility, and vehicular transportation affects the daily lives of most of us. Thus, the ability to forecast the state of traffic in a road network is an important functionality and a challenging task. Traffic data is often obtained from sensors deployed in a road network. Recent proposals on spatial-temporal graph neural networks have achieved great progress at modeling complex spatial-temporal correlations in traffic data, by modeling traffic data as a diffusion process. However, intuitively, traffic data encompasses two different kinds of hidden time series signals, namely the diffusion signals and inherent signals. Unfortunately, nearly all previous works coarsely consider traffic signals entirely as the outcome of the diffusion, while neglecting the inherent signals, which impacts model performance negatively. To improve modeling performance, we propose a novel Decoupled Spatial-Temporal Framework (DSTF) that separates the diffusion and inherent traffic information in a data-driven manner, which encompasses a unique estimation gate and a residual decomposition mechanism. The separated signals can be handled subsequently by the diffusion and inherent modules separately. Further, we propose an instantiation of DSTF, Decoupled Dynamic Spatial-Temporal Graph Neural Network (D2STGNN), that captures spatial-temporal correlations and also features a dynamic graph learning module that targets the learning of the dynamic characteristics of traffic networks. Extensive experiments with four real-world traffic datasets demonstrate that the framework is capable of advancing the state-of-the-art.
1 INTRODUCTION
Traffic forecasting uses sensor histories to predict future road-network conditions, but existing graph-based approaches incompletely model traffic’s spatial-temporal structure. The paper addresses this by separating diffusion and inherent signals and learning dynamic spatial dependencies.
- Traffic forecasting predicts future traffic conditions from historical sensor observations and supports traffic management, urban computing, and public safety.
- Knowledge-driven methods use queuing theory, while data-driven methods range from conventional time-series models to spatial-temporal graph neural networks.
- Traffic signals contain diffusion components from other sensors and inherent components independent of other sensors, but prior studies largely model only diffusion.
- Static adjacency matrices can restrict representation of complex road networks because spatial influence changes across time.
- DSTF separates diffusion and inherent information using residual decomposition and an estimation gate, while D2STGNN adds dynamic graph learning and specialized models.
2 RELATED WORK
STGNNs model traffic as spatial-temporal diffusion processes using graph convolutions and sequential models, but important signal components and dynamic dependencies remain incompletely represented.
- STGNNs construct adjacency matrices for road-network topology and combine diffusion convolutions with sequential models to capture spatial-temporal traffic correlations.
- Recent work improves STGNNs with graph-convolution variants, attention mechanisms, and models targeting long-term temporal dependencies.
- Existing approaches commonly neglect non-diffusion traffic signals and therefore model complex spatial-temporal correlations coarsely.
- Dynamic spatial-dependency methods often use traffic-condition features while ignoring road topology, temporal information, or node information.
3 PRELIMINARIES
The paper represents traffic systems as graphs whose nodes are deployed sensors and defines forecasting as predicting future sensor signals from historical observations.
- A traffic sensor records information such as passing-vehicle flow or vehicle speed within a traffic system.
- A traffic network is a directed or undirected graph G=(V,E), with nodes representing deployed sensors and edges representing network connectivity.
- An adjacency matrix A ∈ R^N×N can represent node reachability using pairwise road-network distances.
- Traffic forecasting maps historical signals from T_h past steps to future signals over T_f nearest future steps.
4 THE DECOUPLED FRAMEWORK
DSTF decomposes traffic signals into diffusion and inherent components in a data-driven manner, then processes them with specialized modules. Its decouple block combines residual decomposition with an estimation gate, while dynamic graph learning supplies adaptive spatial dependencies for diffusion modeling.
- Decoupled Spatial-Temporal Framework: DSTF transforms raw traffic signals into latent representations and decomposes each layer’s input into diffusion and inherent signals.The decomposition is expressed as X = X_dif + X_inh and uses a residual decomposition mechanism plus an estimation gate.
- Decoupled Spatial-Temporal Framework: The diffusion and inherent blocks separately handle the two decomposed hidden time series, allowing each model to focus on its corresponding signal characteristics.Multiple decoupled spatial-temporal layers can be stacked to enhance model capability.
- Dynamic Graph Learning: Dynamic graph learning generates adaptive spatial dependencies for the diffusion model by learning latent correlations between time series.The module uses a self-attention mechanism and adjusts road-network-based spatial dependency using available information.
- Residual Decomposition Mechanism: Residual decomposition removes signal portions approximated by the diffusion or inherent models and retains the residual signals for subsequent processing.Each diffusion and inherent block has a primary model, forecast branch, and backcast branch; the backcast reconstructs learned knowledge for residual removal.
- Estimation Gate: The estimation gate learns values in (0, 1) from node and time embeddings to estimate each traffic series’ diffusion-signal proportion.The gate uses source and target node embeddings together with day and week time-slot embeddings, and broadcasts the estimate across input channels.
- Framework Scope: DSTF leaves dynamic graph learning, diffusion modeling, and inherent modeling abstract so these components can be designed independently for their respective signal characteristics.The paper instantiates the framework by designing these components in the subsequent model.
5 DECOUPLED DYNAMIC ST-GNN
D2STGNN instantiates DSTF by separating diffusion and inherent signals, then models each with specialized spatial-temporal components and learns dynamic spatial dependencies.
- D2STGNN decomposes diffusion and inherent signals so subsequent models can focus on their respective signal types.
- Diffusion Model: The diffusion model uses localized convolution, where recent signals from k_s-order neighbors over k_t past steps influence each target node.Typical k_s and k_t values are 2 or 3.
- Diffusion Model: The localized transition construction masks a target node’s own recent signals, assigning those inherent patterns to the inherent model.
- Diffusion Model: The diffusion module supplements handcrafted road-network transitions with a self-adaptive transition matrix learned from node embeddings.
- Inherent Model: The inherent model combines GRU-based short-term modeling with multi-head self-attention for long-term temporal dependencies.Positional encoding is applied between the GRU and self-attention layers.
- Dynamic Graph Learning: The dynamic graph module learns latent correlations between time series using self-attention to adjust spatial dependencies over time.
12 end
The model is trained by minimizing its loss with gradient descent, using curriculum learning as the training strategy.
- Backpropagation updates the model parameters according to the specified optimization equation.
- The training procedure uses curriculum learning and minimizes the model loss through gradient descent.
6 EXPERIMENTS
Experiments evaluate D2STGNN on four large real-world traffic datasets using established baselines and three standard forecasting metrics.
- Datasets: Experiments use four large real-world datasets containing tens of thousands of time steps and hundreds of sensors.Two datasets contain traffic speed and two contain traffic flow; each has one feature channel.
- Experimental Setup: Traffic networks are constructed from thresholded Gaussian kernels for speed datasets and networks provided by ASTGCN for flow datasets.
- Datasets: METR-LA contains 207 sensors sampled every 5 minutes across 34,272 time slices from March through June 2012.
- Baselines: The study compares D2STGNN with historical, classical regression, recurrent, diffusion-convolutional, and graph-based neural baselines.
- Evaluation: Performance is evaluated with MAE, RMSE, and MAPE, which respectively reflect accuracy, sensitivity to abnormal values, and reduced dependence on data units.
6.2 The Performance of D2STGNN
D2STGNN is evaluated across four traffic datasets and multiple forecasting horizons, with comparisons designed to isolate the effects of decoupling and dynamic graph learning. It consistently achieves the best reported performance, while ablations show that decoupling is important.
- D2STGNN consistently achieves the best performance across all horizons and all four datasets.The comparison includes METR-LA, PEMS-BAY, PEMS04, and PEMS08.
- Final performance reflects temporal and spatial dependency modeling together with dynamic spatial topology, so Table 3 alone cannot isolate the framework’s effectiveness.
- The evaluation removes dynamic spatial dependency learning from all methods to compare decoupled and coupled spatial-temporal designs fairly.The resulting comparisons include D2STGNN variants with static graphs and coupled hidden signals.
- D2STGNN† significantly outperforms the coupled D2STGNN‡, indicating that the decoupled spatial-temporal framework is crucial to the model.
- The coupled D2STGNN‡ still outperforms baselines, but its limited advantage highlights the importance of decoupling the two hidden traffic signals.
6.4 Efficiency
The efficiency study compares average training time per epoch on METR-LA across D2STGNN, its static-graph variant, and several baselines. D2STGNN maintains competitive efficiency while achieving stronger performance than some state-of-the-art methods.
- Average training time per epoch is compared on METR-LA using identical batch size and specified hardware across six models.The compared models are D2STGNN, D2STGNN†, DGCRN, GMAN, MTGNN, and Graph WaveNet.
- D2STGNN does not increase computational burden too much compared with other baselines.
- D2STGNN achieves better performance and higher efficiency than GMAN and DGCRN in the same comparison.
- Graph WaveNet and MTGNN are the most efficient methods, but their performance is worse than that of other models.
6.5 Ablation Study
The ablation study tests the decoupled architecture, its components, and training strategy, while parameter analysis examines spatial, temporal, and hidden dimensions. Results support the importance of decoupling and appropriate model capacity.
- Architecture and components: The estimation gate and residual decomposition mechanism are both important for decoupling the hidden signals.
- Architecture and components: Removing decoupling reduces forecasting accuracy, indicating that separating the two hidden signals is crucial for accurate traffic forecasting.
- Architecture and components: Dynamic graph learning consistently improves performance over a pre-defined static graph.
- Training strategy: Auto-regressive forecasting is more suitable for the model, while curriculum learning helps it converge better.
- Parameter sensitivity: Spatial and temporal kernel sizes from 1 to 5 are tested, supporting localized diffusion characteristics.A smaller hidden dimension is insufficient to encode spatial-temporal information, whereas a larger one may introduce overfitting.
6.7 Visualization
The visualization compares predictions with ground truth at two randomly selected sensors in the METR-LA test set over June 13–16, 2012. The two sensors exhibit different traffic patterns.
- Predictions and ground truth are visualized for nodes 2 and 111 over four days from the METR-LA test dataset.
- The selected sensors display different traffic patterns, including differing congestion behavior.
7 CONCLUSION
The paper proposes DSTF to separate diffusion and inherent traffic signals, and instantiates it as D2STGNN with specialized modeling and dynamic graph learning. Experiments on four real-world datasets show consistent, significant improvement over all baselines.
- DSTF decouples diffusion and inherent signals from traffic data for more precise modeling.
- D2STGNN models diffusion signals with spatial-temporal localized convolution and inherent signals with recurrent neural networks plus self-attention.
- Its dynamic graph learning module adjusts road-network spatial dependencies by learning latent correlations between time series.
- Four real-world datasets show that D2STGNN consistently and significantly outperforms all baselines.