WorksComputer Science
SIMBAD: Spatio-Temporal Traffic Forecasting Robust to Aperiodicity
Abstract
Accurate traffic prediction is essential for urban planning but remains challenging due to the irregular patterns caused by unexpected events. While Spatio-Temporal Graph Neural Networks (STGNNs) effectively model periodic data, they often struggle with these aperiodic fluctuations. To address this, we propose SIMBAD, which improves generalization irregular patterns via two key mechanisms: 1) adaptive regulation of periodic signals to prioritize recent signals, and 2) dynamic spatial modeling that adjusts influence between connected nodes based on contextual relations. Built upon simple similarity metrics to adapt to aperiodic patterns, SIMBAD outperforms existing methods on real-world benchmark datasets, demonstrating superior performance in predicting unseen and irregular traffic events.
Introduction
Accurate spatio-temporal forecasting is a prominent research area due to its broad applicability, particularly in traffic flow forecasting [9] for Intelligent Transportation Systems (ITS) [36]. Despite its practical importance, accurate spatio-temporal traffic forecasting remains challenging due to inherent complex dependencies and irregularities. As shown in Figure 1(a), traffic data from interconnected roads often exhibits heterogeneous patterns, such as temporal variations between periodic intervals (e.g., sensor 137) and spatial variations between neighboring nodes (e.g., sensors 123 and 137). To address this aperiodicity and achieve accurate forecasting, recent studies [14, 20] have utilized semantic adjacency matrices based on Dynamic Time Warping (DTW) [2]. However, the high computational complexity of DTW limits its scalability, motivating the use of more efficient alternatives like Minkowski distance [29]. The use of historical information is grounded in extensive empirical evidence on the recurrent nature of human mobility and urban traffic. Prior studies have shown that human mobility exhibits highly reproducible spatial and temporal patterns [12, 31], while urban congestion exhibit substantial day-to-day regularity [25]. Notably, as shown in Figure 1(b), we observe that the simple distance between current and historical data (e.g., last week) is still sufficient to show a strong correlation between the corresponding distance of the input sequence and that of the target sequence. It suggests that lightweight distance metrics can effectively quantify underlying regularities in traffic patterns.

Figure 1. Motivating examples on PEMS04 dataset. (a) Varying periodic traffic patterns among connected nodes. (b) Correlation between L1-norms of input and target sequences with respect to their last week counterparts. (c) Performance comparison under growing aperiodicity.
Besides, another line of research leverages the strong periodicity of traffic data, such as recurring peak hours on weekdays. For instance, ST-ResNet [39] explicitly models temporal closeness, period and trend, while ASTGCN [13] incorporates weekly, daily, and recent sequences to better predict periodic patterns. Subsequent works [4, 23] further advanced this by introducing periodic positional encodings, enabling models to represent recurring temporal regularities at each position more effectively. Despite these successes, reliance on historical sequences can degrade performance. Revisiting Figure 1(a), traffic patterns often vary even on the same day across different weeks. This discrepancy at a certain timestep tends to propagate to its future states as empirically quantified in Figure 1(b), making such aperiodic historical sequences less informative for predicting the target sequences. Consequently, existing models that overlook these variations experience significant performance degradation as the aperiodicity of the target sequence increases.
To address these challenges, we propose SIMBAD, a simple yet effective SIMilarity-Based Aperiodicity-robust forecasting methoD. We explicitly exploit similarity between historical and input traffic patterns, enabling adaptive learning of spatio-temporal dependencies. Specifically, we demonstrate that the lightweight distance between current and past periodic data can sufficiently guide a graph neural network, enabling SIMBAD to adaptively consult past traffic information and capture spatio-temporal patterns. In addition, we adopt the uncertainty measure [16] to regulate noise in the distance when quantifying the discrepancy observed between current and past traffic patterns. SIMBAD then regulates information flow by modulating spatial interactions between nodes and selectively incorporating historical data that aligns with recent dynamics, ensuring robust forecasting under aperiodic conditions. Consequently, SIMBAD is particularly more effective in forecasting aperiodic patterns under unexpected sporadic events. As shown in Figure 1(c), SIMBAD exhibits significantly lower performance degradation compared to the seminal periodicity-aware method [13] as the aperiodicity of the target sequence increases.
In brief, our contributions are summarized as follows:
We propose a simple yet effective similarity-based spatio-temporal forecasting approach, overcoming the practical challenges of handling sporadic, irregular patterns.
We present a graph neural network framework SIMBAD, employing an Aperiodicity-Aware Modulation Layer and a Spatio-Temporal Modeling layer, to dynamically adjust the flow of temporal information between connected nodes. Furthermore, incorporating the proposed Aperiodicity-Aware Modulation Layer into existing baselines improves their predictive performance, demonstrating its effectiveness as a transferable module. The source code is publicly available at
https://github.com/thegk3/SIMBAD.Extensive experiments on four widely used benchmark datasets, and three supplementary datasets with different regions or large-scale, demonstrate that our model not only achieves strong overall performance but also excels at predicting unseen, irregular traffic patterns, achieving 4.12% of accuracy improvement on average over the best baseline.
Preliminary
We formalize the prediction problem in a traffic network and the notations used throughout the paper.
Definition 2.1. The traffic network is represented as a graph , where is a set of nodes, is a set of edges, and is the weighted adjacency matrix of graph .
Given previous traffic observation of timesteps, the typical goal of traffic forecasting is to predict future traffic flow of timesteps with input by learning a mapping function :
Definition 2.2. Given a timestep , we define two auxiliary traffic windows of length , which represent the preceding and subsequent context window of , respectively:
Then, we introduce the two types of inputs at a given timestep :
Recent input represents the previous observation of timesteps preceding :
Periodic input represents historical windows of length starting at various periodic intervals prior to the onset of the target sequence:
where is a set of periodic intervals. For example, with days and the number of timesteps in a day, and corresponds to the historical window observed days prior to the target sequence.
Given these inputs, we aim to learn the mapping function which maps the multiscaled historical information to future traffic states:
SIMBAD Framework
Figure 2 shows the overall framework of SIMBAD with two key layers. The aperiodicity-aware modulation layer exploits temporal and spatial similarities to adaptively modulate historical inputs with uncertainty and semantic spatial dependencies; the spatio-temporal modeling layer models the regulated temporal and spatial signals to reflect spatio-temporal dynamics in target prediction.

Figure 2. (a) Input segments of SIMBAD. (b) Main components of SIMBAD. Overall framework of SIMBAD.
Aperiodicity-Aware Modulation Layer
Similarity-driven Temporal Regulation with Uncertainty
SIMBAD regulates the periodic inputs based on the similarity with the recent input by -norm:
where and represent two different traffic windows.
Specifically, we quantify the relevance of periodic input to the current state based on the similarity of the current traffic sequence and the target sequence to each corresponding past periodic sequences, respectively:
To address the discrepancy between and , we model the aleatoric uncertainty caused by noise inherent in the similarity measurement inspired by the prior work [16]. We first encode into a latent space and train a lightweight predictor and uncertainty head to estimate and the log variance , respectively. The formulated uncertainty enables the model to be more robust to noisy data by mitigating the effect of erroneous signals. Beyond robust similarity estimation, the predicted similarity and uncertainty are explicitly incorporated into the model to modulate how periodic input is utilized.
Uncertainty-Guided Aggregation. We estimate a similarity-based probability distribution to adaptively aggregate multiple periodic inputs based on the predicted similarity . Specifically, we incorporate the uncertainty as a confidence signal to down-weight the uncertain similarity signal:
where is a sigmoid function and is a temperature. Then, an aggregated periodic input is estimated:
SIMBAD further constructs a masking matrix to regulate the inclusion of the historical information. Since the periodic input is aggregated under the probability distribution in Eq. (8), we expect the input to follow the estimated similarity and confidence as:
Then, we construct a mask with a learnable threshold :
We then multiply the constructed masking matrix by the aggregated input to regulate the inclusion of of each node and obtain the final historical input:
where is the Hadamard product.
In this way, SIMBAD dynamically controls periodic input selection and feature masking, while the uncertainty makes it remain robust under irregular and noisy traffic conditions.
Similarity-driven Semantic Spatial Regulation
In a traffic network, the connectivity information between sensors plays a critical role in modeling information propagation from a spatial perspective. Prior works [13, 21, 37] model the spatial information using Chebyshev spectral filter [7] or incorporate Gaussian kernel-based adjacency [28], which assumes that spatial correlations decay with distance. However, in real-world traffic scenarios, fixed states such as simple connections or distances do not necessarily reflect the true spatial dependencies among sensors, as they can vary depending on traffic conditions and temporal context. To address this limitation, we construct similarity-driven semantic adjacency matrices that can dynamically regulate spatial information.
Self-adaptive Adjacency Matrix. An actual connectivity between nodes may not necessarily lead to their spatial dependency. Since the historical input in Eq. (12) represents the most contextually relevant periodic state, it is important to group nodes that exhibit similar periodic behaviors. Thus, we define semantically meaningful dependencies, rather than relying on fixed connections.
Past works [21, 35] proposed a self-adaptive adjacency matrix to guide the model to discover hidden spatial dependencies by randomly initializing node embedding dictionaries. Instead of random initialization, we initialize the node embeddings based on the reliability of historical information using and . Specifically, we initialize each node according to its similarity and confidence to the current state as . Then, with learnable parameters , two node embeddings are initialized as:
Finally, the self-adaptive adjacency matrix is constructed as:
where the ReLU activation function eliminates weak connections, and the Softmax operation normalizes the self-adaptive adjacency matrix. In this way, the model is able to learn semantically meaningful connections based on the reflection of historical information to current traffic information.
Similarity-based Adjacency Matrix. Since the prediction target corresponds to the immediate future following the current state, the actual connectivity reflected in the recent input plays a crucial role in determining how information propagates across nodes, encouraging the need to use real connectivity relations. However, relying solely on connectivity information does not model the potential variations in information among nodes. We further weight the pairwise node connections dynamically depending on their recent input similarities based on the L1 distance:
where denotes the L1-distance between nodes and .
Since this spatial similarity is derived from the same sequence representation of temporal similarity in Eq. (7), we weight the spatial connections using the same threshold in Eq. (11) to maintain consistent interpretations of similarity across temporal and spatial modeling. Thus, given the predefined adjacency matrix , we weight each connection between node and as:
Finally, we normalize the adjacency matrix for learning stability:
Spatio-Temporal Modeling Layer
A spatio-temporal modeling layer contains three key components: 1) Self-Attention Block where vanilla transformers capture both spatial and temporal correlations, 2) Spatial Convolution Block which handles spatial dependencies from -hop neighbors with various aggregation methods, and 3) Gated Temporal Convolution Block which focuses on catching temporal dependencies from nearby timesteps in a node. To effectively process the hidden periodic information, first layers process the historical data, followed by layers which process the recent data. Let hidden dimension be .
Self-Attention Block. We apply vanilla transformers to capture traffic relations in various perspectives. Given the input , we apply self-attention to capture dependencies within data. We first compute the query (), key () and value () matrices as:
where , , are learnable parameters. Then, we compute the attention score matrix and apply the attention score to the value matrix as . We apply this calculation along spatial and temporal axes separately. Thus, each attention scores are and which contains spatial relations of different timesteps and temporal relations for different nodes, respectively. The output of each attention block is defined as:
Spatial Convolution Block. Graph convolution is an essential operation in extracting a node’s features within the structural information. Studies [3, 7] proposed a graph convolutional method based on spectral graph theory and [17] proposed a first approximation of Chebyshev spectral filtering. These spectral-based approaches enable the modeling of spatial dependencies by performing convolution in the graph Fourier domain. A single-hop graph convolution is defined with the normalized adjacency matrix , input , and learnable parameter :
Then, the spatial convolution block is defined to aggregate information from -hop neighbors:
Gated Temporal Convolution Block. In addition to capturing local dependencies, it is crucial to model temporal dynamics by merging information across neighboring timesteps. To address this, we apply 1D convolution along the temporal axis to learn local temporal patterns. However, standard 1D convolution may treat all timesteps equally and fail to distinguish between irregular timesteps. To overcome this limitation we incorporate a Gated Tanh unit (GTU) [6] to capture the temporal dynamic information of traffic flow data. GTU in time dimension can be defined as:
where and are two independent 1D convolution operations.
Optimization
Given input and adjacency matrix , the output of SIMBAD layer is obtained by stacking the temporal modeling, then spatial modeling, respectively:
As mentioned, first layers process the historical data to obtain . Since the dissimilar nodes of historical input are masked as in Eq. (12), the recent input of corresponding nodes must be emphasized for accurate prediction. Thus, we add the whole recent input and the recent input of corresponding nodes through :
Then, we generate the final output
where is the horizon of prediction. In our case, the output dimension is 1. We supervise the uncertainty derived from discrepancy between equations in Eq. (7) [16]:
To predict the future timesteps , we minimize the loss:
Finally, SIMBAD aims to jointly minimize the combined loss with a regularization factor :
Experiments
Experiment Settings
Datasets and Evaluation Metrics We verify the performance of SIMBAD mainly on four widely used benchmark datasets [30], PEMS03, PEMS04, PEMS07, and PEMS08. Traffic flow is measured every 5 minutes, resulting in 288 timesteps in a day, and standard normalization is applied to the input data. Details of datasets are given in Table 1.
| DATASET | NODES | EDGES | TIMESTEPS |
| PEMS03 | 358 | 547 | 26,208 |
| PEMS04 | 307 | 340 | 16,992 |
| PEMS07 | 883 | 866 | 28,224 |
| PEMS08 | 170 | 295 | 17,856 |
Table 1. Summary of datasets.
We also employed three supplementary datasets (METR-LA, PEMS-BAY, and LargeST) with the evaluation results provided in Appendix § 14. We used the three most frequently used evaluation metrics in traffic forecasting: root mean squared error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). We calculate the error based on the average performance of all predicted horizons on the test sets.
Baselines We compared SIMBAD with 11 baselines categorized into three groups: (1) GNN-based (GWNet [35], DCRNN [21], AGCRN [1], STSGCN [30], DSTAGNN [18], and ASTGCN [13]), (2) Transformer-based (STAEformer [23], STWave [10], DTRformer [5], and LHMformer [26]), and (3) Memory-based (STIDGCN [22]). More details are given in Appendix § 15.
Implementation Details To align with recent studies, we split datasets into train, validation, and test sets in a ratio. We train our model using Adam optimizer with a learning rate of 0.001 and the batch size of 64. We use 2 weeks, 1 day, and 1 hour of historical data for the prediction of the next 12 timesteps. The hidden dimension is 64, and the number of layers and are both set to 2. Temperature is set as 0.2 and regularization factor is 0.1. We use 4 heads for self-attention transformers. The training epoch is set to 100 with an early-stop approach for which the validation error does not improve within 30 epochs. All baselines used the default hyperparameters suggested in the original work or were fine-tuned to achieve the best scores.
Overall Performance Results
Table 2 shows the results of eleven baselines, SIMBAD and ablation results. SIMBAD outperforms baseline models in most cases, demonstrating its overall prediction capability. Notably, SIMBAD demonstrates superior performance not only on small-scale datasets such as PEMS08, but also on PEMS07 with large and complex network structures. Specifically, SIMBAD achieved an average performance gain of 5.80% and 6.66% on PEMS08 and PEMS07, respectively, over the strongest baseline.
| Model | PEMS03 | PEMS04 | PEMS07 | PEMS08 | ||||||||
| Metric | MAE | RMSE | MAPE% | MAE | RMSE | MAPE% | MAE | RMSE | MAPE% | MAE | RMSE | MAPE% |
| GWNET | 14.94 | 26.06 | 15.52 | 19.23 | 30.76 | 14.04 | 20.63 | 33.26 | 9.14 | 15.06 | 23.79 | 9.73 |
| DCRNN | 16.67 | 29.06 | 16.60 | 21.68 | 34.19 | 15.05 | 23.13 | 36.13 | 10.18 | 16.81 | 25.91 | 11.71 |
| AGCRN | 16.21 | 28.57 | 15.35 | 19.85 | 32.96 | 13.38 | 21.35 | 35.12 | 9.02 | 16.54 | 26.23 | 10.47 |
| STSGCN | 18.24 | 30.51 | 17.37 | 21.46 | 34.07 | 14.37 | 24.07 | 38.69 | 10.17 | 17.05 | 26.47 | 10.87 |
| STAEformer | 16.67 | 29.56 | 17.21 | 18.63 | 30.86 | 13.10 | 19.91 | 33.94 | 8.58 | 13.82 | 23.55 | 9.23 |
| DSTAGNN | 15.99 | 28.20 | 15.45 | 19.85 | 31.91 | 13.42 | 21.84 | 34.95 | 9.58 | 16.19 | 25.44 | 10.61 |
| STWave | 15.91 | 28.06 | 16.62 | 19.63 | 32.02 | 13.66 | 21.32 | 35.68 | 8.99 | 14.01 | 23.94 | 9.33 |
| STIDGCN | 15.98 | 26.66 | 17.24 | 18.38 | 30.47 | 12.42 | 19.20 | 32.87 | 8.18 | 13.71 | 23.33 | 9.00 |
| DTRformer | 15.83 | 27.61 | 16.31 | 18.49 | 30.62 | 13.08 | 19.52 | 32.53 | 8.38 | 13.50 | 22.90 | 9.05 |
| LHMformer | 16.22 | 27.81 | 16.76 | 18.45 | 30.70 | 13.00 | 19.27 | 32.54 | 8.11 | 13.66 | 23.04 | 9.02 |
| ASTGCN | 20.25 | 35.99 | 20.67 | 22.20 | 35.24 | 15.66 | 25.50 | 39.75 | 11.43 | 17.07 | 26.49 | 11.80 |
| SIMBAD | 14.77 | 25.87 | 14.35 | 18.03 | 30.08 | 12.36 | 17.86 | 30.99 | 7.44 | 12.58 | 22.13 | 8.35 |
| w/o | 15.94 | 28.42 | 16.65 | 18.48 | 30.74 | 12.96 | 18.46 | 31.80 | 7.91 | 13.10 | 24.21 | 9.07 |
| w/o | 15.19 | 26.99 | 15.25 | 18.25 | 30.32 | 12.60 | 18.23 | 31.40 | 7.69 | 12.91 | 22.31 | 8.79 |
| w/o A | 15.65 | 27.79 | 16.56 | 18.26 | 30.14 | 12.91 | 18.02 | 31.05 | 7.63 | 12.72 | 22.14 | 8.70 |
Table 2. Performance comparison and ablation study results on PEMS datasets.
We conduct a series of ablation experiments on SIMBAD. (1) : remove the masking matrix on the input, using every historical data. (2) : remove uncertainty and weight historical input based on Eq. (7). (3) w/o A: use unweighted adjacency matrices for all input. Table 2 shows the effectiveness of different components in our SIMBAD model. For all datasets, removing the masking matrix , which regulates historical inputs based on similarity to current traffic states, leads to the largest performance degradation, highlighting the crucial role of regulating historical inputs. Also, a significant drop in performance is observed by the removal of dynamic weights among nodes (w/o A), indicating that regulating dissimilar patterns among connected nodes is also essential for robust spatio-temporal forecasting. Furthermore, the removal of calculating uncertainty also causes a substantial decrease in performance, showing the necessity of regularizing misleading signals.
Performance Analysis with Increasing Irregularity
Figure 3(a) presents the trend of prediction accuracy of all datasets with respect to increasing -norm value of target timestep to the past week. We compared SIMBAD to four types of baselines: (1) ASTGCN [13] which leverages historical data for prediction; (2) GNN-based model DSTAGNN [18]; (3) a Transformer-based LHMformer [26]; (4) memory-based STIDGCN [22]. SIMBAD consistently outperforms other baselines as the -norm of the target sequence and its corresponding historical data increases, while maintaining the superior overall performance. In particular, the largest performance gain over ASTGCN on all datasets indicates that the regulation of irregular patterns in historical datasets is crucial, especially in forecasting aperiodic patterns. Also, SIMBAD demonstrates its strong robustness to growing aperiodicity throughout all datasets. For instance, in PEMS03, DSTAGNN, LHMformer, STIDGCN exhibit comparable performance across the whole calculated -norm range. However, DSTAGNN and LHMformer suffer from substantial performance degradation in PEMS04 and STIDGCN achieves only the third-best performance in PEMS07. These results demonstrate that SIMBAD can effectively distinguish and maintain its robustness to sporadic, irregular patterns while preserving overall performance.

Figure 3. (a) shows performance comparison over increasing irregularity. (b) shows an example prediction result of SIMBAD and the best recent baseline LHMformer from Node 88 of PEMS08.
Figure 3(b) presents an example of performance comparison between LHMformer [26] and SIMBAD under an irregular traffic scenario where the current traffic context significantly differs from both the 1 week and 1 day historical patterns. Since LHMformer [26] leverages historical information using a full-day window to emphasize daily periodicity, its prediction becomes strongly influenced by the daily historical pattern, resulting in performance degradation when such patterns fail to generalize to the current context. Additional case study results are provided in Appendix § 8.§ 2.
Hyperparameter Sensitivity Analysis
We further investigate hyperparameter sensitivity on the tested datasets. We select two hyperparameters for investigation: (1) temperature , which controls the sharpness of the distribution in Eq. (8); (2) regularization factor , which regularizes the loss in Eq. (26). The default values for and are 0.2 and 0.1, respectively, and the results are shown in Figure 4.

Figure 4. Hyperparameter sensitivity analysis in terms of MAE on four datasets.
Results demonstrate that SIMBAD maintains stable performance across a wide range of hyperparameter configurations. Furthermore, performance degradation can be observed with a high temperature value. A high value leads to a smoother distribution in Eq. (8) and reduces the distinction between reliable and unreliable periodic inputs. This causes the dissimilar historical patterns to be more likely to be aggregated, weakening the similarity-based periodic selection process and leading to performance degradation. Also, as the regularization factor increases, the uncertainty regularization term exerts a stronger influence on the optimization process. Excessive regularization may overemphasize uncertainty minimization rather than forecasting accuracy, which can suppress useful periodic information and degrade the overall prediction performance.
Modularity of SIMBAD
To evaluate its applicability beyond SIMBAD, we integrate the temporal regulation mechanism in § 3.1.1 into existing forecasting architectures. The module regulates periodic inputs based on their relevance and uncertainty before passing them to the forecasting backbone, allowing each baseline’s original forecasting architectures to be preserved. This modular design enables existing architectures to selectively exploit historical information without redesigning their core components.
As shown in Table 3, the module reduces MAE in 9 of 12 cases when equipped with three representative baselines. DSTAGNN and STIDGCN achieve average relative reductions of 2.51% and 2.93%, respectively, with improvements reaching 8.03% and 4.30% on PEMS08. LHMformer achieves reductions of 0.43% on PEMS03 and 0.95% on PEMS08. These results demonstrate that regulating historical information according to its relevance and uncertainty can also enhance existing forecasting architecture. This confirms that the proposed temporal regulation mechanism can serve as a general and modular component for improving robustness to irregular historical patterns. Additional results on other baselines and metrics are provided in Appendix B.
| Model | PEMS03 | PEMS04 | PEMS07 | PEMS08 |
| DSTAGNN | 15.99 | 19.85 | 21.84 | 16.19 |
| + Ours | 16.44 | 19.53 | 21.14 | 14.89 |
| STIDGCN | 15.98 | 18.38 | 19.20 | 13.71 |
| + Ours | 15.20 | 18.27 | 18.83 | 13.12 |
| LHMformer | 16.22 | 18.45 | 19.27 | 13.66 |
| + Ours | 16.15 | 18.83 | 19.76 | 13.53 |
Table 3. Modularization results in terms of MAE.
Historical Window Analysis
In this section, we analyze the contribution of each historical window. In our settings, we used 2 weeks, 1 week and 1 day periodic data for prediction. We show results when the weekly historical data and daily historical data are not considered for calculation, respectively. Results are shown in Table 4. As shown in the table, the historical information is indeed important for accurate prediction. In particular, removing the weekly historical data leads to an average MAE increase of 9.27% across all datasets, compared to only 1.86% when the daily historical data is removed. The most pronounced degradation is observed on PEMS08, where removing the weekly historical data increases MAE by 15.66%, compared to 2.94% when the daily historical data is removed. These results suggest that weekly historical information plays a more critical role in accurate traffic prediction. This is a valid result, considering that traffic data exhibits strong weekly periodicity, where mobility patterns are largely governed by recurring schedules such as weekday commuting and weekend behavior.
| Model / Metric | PEMS03 | PEMS04 | PEMS07 | PEMS08 | ||||||||
| MAE | RMSE | MAPE% | MAE | RMSE | MAPE% | MAE | RMSE | MAPE% | MAE | RMSE | MAPE% | |
| SIMBAD | 14.77 | 25.87 | 14.35 | 18.03 | 30.08 | 12.36 | 17.86 | 30.99 | 7.44 | 12.58 | 22.13 | 8.35 |
| No week | 15.29 | 26.79 | 16.14 | 19.54 | 31.59 | 13.19 | 19.56 | 32.93 | 8.32 | 14.55 | 23.52 | 9.41 |
| No day | 15.02 | 25.91 | 14.55 | 18.30 | 30.40 | 12.56 | 18.09 | 31.35 | 7.50 | 12.95 | 22.53 | 8.68 |
Table 4. Performance comparison on different historical windows.
Related Works
Traditional time series forecasting methods, such as ARIMA [34], VAR [41] and SVR [8], focused on modeling temporal dependencies, and recurrent architectures, such as RNNs and LSTMs, were introduced to better capture temporal context. Then, traffic forecasting studies [13, 30, 37] incorporated spatial modeling through graph-based approaches, enabling representation of interdependencies among road sensors and capturing spatial information.
Spatio-Temporal Traffic Forecasting. Based on spatial relations between sensors, traffic forecasting requires handling spatial correlations between data, different from general time series prediction. To effectively model the traffic flow based on both spatial and temporal dependencies, various studies have been conducted. DCRNN [21] models the traffic flow as a diffusion process, STGCN [37] incorporates 1D temporal convolution along with spatial graph convolution to capture spatio-temporal dependencies. Models such as ASTGCN [13], STAEformer [23], LHMformer [26] incorporated attention mechanisms which achieved promising performance. PDFormer [14], GMAN [40], STID [27] applied spatial embedding to capture inherent spatial dependencies. To capture the temporal and spatial dependencies simultaneously, STSGCN [30] modeled the spatial and temporal blocks altogether. STFGNN [20] and PDFormer [14] implemented DTW [2] to measure the similarity between nodes and used them to build a similarity-based temporal graph. Recent models, such as PatchSTG [11], achieved great performance by combining Transformer with a tree structure, but it requires auxiliary information such as latitude and longitude.
Memory-Based Traffic Forecasting. With the success of memory networks in natural language processing [33], various traffic forecasting studies also adopted the idea of memory networks to store and utilize repeating traffic patterns. For instance, PM-MemNet [19] proposes that traffic data can be summarized into numerous clusters of representative patterns by using pre-calculated patterns. It matches input data to representative patterns with a key-value structure and uses the gained information from the memory for traffic forecasting. MNNCCRN [32] combines the concepts of memory network and cluster to capture low- and high-level patterns more accurately. Also, MegaCRN [15] leverages the idea of memory networks by building a Meta-Node Bank. Based on this bank, MegaCRN generates an adaptive meta-graph that is robust and adaptive to normal and non-stationary traffic situations. These memory-based approaches require the employment of a dedicated memory component for exploiting traffic patterns in addition to the traffic forecasting model, whereas SIMBAD exploits only the similarity-based proxy derived from the input sequences to adaptively control the periodic traffic patterns.
Conclusion
In this work, we introduce SIMBAD, a novel spatio-temporal traffic forecasting framework to enhance prediction performance, particularly under aperiodicity. By adaptively balancing the influence of periodic temporal patterns and spatial dependencies through similarity-aware weighting, SIMBAD addresses the limitations of existing approaches when confronted with unexpected traffic fluctuations observed in historical periodic inputs. SIMBAD exploits simple and lightweight similarity-based weights through an aperiodicity-aware modulation layer and a spatio-temporal modeling layer to refine the flow of temporal information across graph nodes, leading to more robust and accurate forecasting. Comprehensive experiments on seven real-world datasets validate the effectiveness of SIMBAD, demonstrating its superior performance in both regular and irregular traffic scenarios. Furthermore, the modularity of SIMBAD has been demonstrated by showing that adopting the temporal regulation mechanism in existing baselines can improve their performance. These results underline the potential of exploiting simple similarity-based historical signals in advancing Intelligent Transportation Systems and supporting more reliable traffic management.
Acknowledgments and Disclosure of Funding
This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grants (No. IITP-2026-RS-2025-02304828(40%), No. IITP-2026-RS-2024-00436857(20%), No. RS-2026-25507543(20%), No. IITP-2026-RS-2020-II201819(10%)) and the National Research Foundation of Korea (NRF) grant (No. RS-2025-25435830(10%)) funded by the Korea government (MSIT).
Case Study
Effect of Uncertainty
In this section, we further analyze the effect of uncertainty we proposed. The uncertainty reflects the amount of discrepancy between and in Eq. (7), where high value means high discrepancy. Therefore, as in Eq. (8), we weight the predicted similarity to reduce the contribution of historical instances whose input context conflicts with the target. To validate this behavior, we visualize the effect of uncertainty on datasets PEMS04 and PEMS08 to observe whether uncertainty properly reduces the contribution of historical instances with high discrepancy.
The left-side of Figure 5 shows that correlation of L1-norm is mostly positive, implying that the aggregated historical input generally follows the traffic flow tendency to that of the target. However, there exist sections where the L1-norm with is low, whereas the L1-norm with is high. As illustrated in the right panels, uncertainty assigns the mismatched sections as uncertain, thus reducing the contribution of the unreliable inputs as in Eq. (8). This demonstrates that the proposed uncertainty mechanism successfully identifies and suppresses the misleading historical patterns.

Figure 5. (a) PEMS04. (b) PEMS08. For each dataset, the left figure shows correlation between L1-norm of target and aggregated historical input in Eq. (9) to L1-norm of target and the periodic inputs in Eq. (4). The right figure shows the uncertainty value for each L1-norm correlation. Higher value means higher uncertainty.
Also, there is an extra interesting observation that the lower-right region of the plots, where the L1-norm with is high but the L1-norm with the target is low, is largely absent. Notably, as shown in Figure 1b, data points in such region are observed. However, our aggregated input does not exhibit this phenomenon, demonstrating that our model does not mistakenly mask the highly-correlated situations.
Irregular Pattern Prediction
In this section, we evaluate the model performance on irregular timesteps, where the target traffic patterns deviate significantly from historical trends. We mainly focus on representative cases where both weekly and daily patterns differ from the target, as traffic data generally follows strong temporal periodicity, where the same day of the week shows similar patterns and recent daily observations capture daily recurring trends.
As illustrated in Figure 6, baseline models inherit dominant historical trends, leading to systematic errors when these patterns fail to generalize to the current context. For example, ASTGCN [13] directly utilizes weekly and daily historical inputs without considering their reliability, leading to incorrect prediction of the evening peak. As shown in Figure 6a, the discrepancy between the weekly and daily evening peak patterns introduces conflicting signals, which degrades the prediction accuracy. Also, LHMformer [26] leverages the historical data with a full-day window, which emphasizes recent daily patterns. As a result, as shown in Figure 6d, the predicted magnitude of the evening peak tends to follow that of the previous day.

(a) PEMS03 – Node 209.

(b) PEMS04 – Node 247.

(c) PEMS07 – Node 674.

(d) PEMS08 – Node 88.
Figure 6. Left-side figure shows the current and the respective daily and weekly historical trends. Middle figure shows the prediction performance of comparing baselines. Right-side figure shows the prediction performance of SIMBAD and SIMBAD without masking matrix of Eq. (11).
In contrast, our model addresses these limitations through an uncertainty-aware aggregation mechanism, which evaluates both the relevance and reliability of historical inputs. This allows the model to suppress misleading patterns and accurately capture the true traffic dynamics, even when both weekly and daily references are inconsistent.
We also evaluate the influence of masking matrix in Eq. (11), which filters out dissimilar labels based on the calculated similarity between historical data. Results show that the masking matrix is critical for correctly predicting the irregular traffic patterns. For example, in Figure 6b, SIMBAD without masking matrix tends to predict the morning peaks to follow as in previous day and week. Also, in Figure 6c, SIMBAD without mask also predicts the morning peaks similar to the previous day.
Detection of Irregularity
In this section, we visualize how SIMBAD detects irregular traffic patterns through its internal variables: the predicted similarity , uncertainty and masking matrix . We also investigate whether the performance improvement is concentrated in such irregular situations. Figure 7 presents two representative cases. In periodic regimes, SIMBAD accurately predicts the similarity between the current pattern and each periodic reference. When traffic becomes aperiodic, the predicted similarities may deviate from their ground-truth values across all periodic references. In these cases, the model assigns higher uncertainty, resulting in low mask values that suppress unreliable periodic information. Accordingly, the largest error reductions over the variant without occur in the highlighted aperiodic intervals, demonstrating the effectiveness of uncertainty-aware masking under irregular traffic conditions.

Figure 7. From top to bottom, the panels show the traffic forecasts, predicted similarities and their ground-truth values , uncertainties , value from masking matrix , and the error reduction over the variant without . Here, the subscripts 2W, 1W, and 1D denote the periodic reference patterns from two weeks, one week, and one day earlier, respectively. Blue shading marks intervals where all exceed the 90th percentile and .
Analysis on Uncertainty
In this section, we clarify and justify the use of uncertainty to model the discrepancy between the -norm of current data and historical data and the -norm of target data and its corresponding historical data. We adopt aleatoric uncertainty from [16] to model such discrepancy. In [16], aleatoric uncertainty is introduced to capture the inherent stochastic noise in observations that cannot be completely eliminated even with sufficient data. Following this perspective, we interpret the discrepancy between the input-historical -norm and the target-historical -norm as an intrinsic stochastic variability in historical similarity estimation.
Figure 8 illustrates the discrepancy distributions across all datasets. The distributions are consistently centered around zero while exhibiting non-negligible variance, indicating that historical similarity generally provides meaningful information but is not perfectly deterministic. Even when the current input exhibits periodic patterns similar to historical observations, the target may deviate from the corresponding historical behavior due to irregular traffic dynamics and stochastic temporal variations.

Figure 8. (a) PEMS03. (b) PEMS04. (c) PEMS07. (d) PEMS08. Figure 8: -norm discrepancy distribution.
Furthermore, the discrepancy distributions exhibit approximately Gaussian-like shapes centered around zero, which is consistent with the probabilistic assumption commonly adopted in aleatoric uncertainty modeling [16]. Therefore, the discrepancy can be interpreted as an input-dependent stochastic noise arising from the imperfect correspondence between current-historical similarity and target-historical similarity. Based on this interpretation, the proposed uncertainty module estimates the reliability of historical similarity information, where larger uncertainty indicates that the historical similarity relationship is less reliable for forecasting the future target.
Modularity of SIMBAD
In this section, we show the full results of modularization of SIMBAD discussed in § 4.5 on all baselines and metrics. Table 5 below shows results of baseline models with and without the proposed Aperiodicity-Aware Modulation Layer. Among the 120 results, the proposed module improves 77 cases, yielding an average relative improvement of 2.00%. Seven of the ten backbones achieve positive average gains, with substantial improvements in DCRNN [21] and STSGCN [30] with average error reductions of 7.97% and 7.87%, respectively. PEMS07 and PEMS08 gain largest average improvements, with an average of 4.71% and 3.78%, respectively. These results demonstrate that the proposed Aperiodicity-Aware Modulation Layer can effectively enhance diverse spatio-temporal forecasting architectures.
| Model | PEMS03 | PEMS04 | PEMS07 | PEMS08 | ||||||||
| Metric | MAE | RMSE | MAPE% | MAE | RMSE | MAPE% | MAE | RMSE | MAPE% | MAE | RMSE | MAPE% |
| GWNET | 14.94 | 26.06 | 15.52 | 19.23 | 30.76 | 14.04 | 20.63 | 33.26 | 9.14 | 15.06 | 23.79 | 9.73 |
| + Ours | 15.96 ▼ | 29.15 ▼ | 16.55 ▼ | 18.80 ▲ | 30.68 ▲ | 13.17 ▲ | 19.11 ▲ | 31.71 ▲ | 8.40 ▲ | 14.10 ▲ | 22.91 ▲ | 9.61 ▲ |
| DCRNN | 16.67 | 29.06 | 16.60 | 21.68 | 34.19 | 15.05 | 23.13 | 36.13 | 10.18 | 16.81 | 25.91 | 11.71 |
| + Ours | 15.91 ▲ | 27.80 ▲ | 16.51 ▲ | 20.12 ▲ | 32.18 ▲ | 13.75 ▲ | 19.99 ▲ | 32.68 ▲ | 8.55 ▲ | 15.62 ▲ | 24.70 ▲ | 10.11 ▲ |
| AGCRN | 16.21 | 28.57 | 15.35 | 19.85 | 32.96 | 13.38 | 21.35 | 35.12 | 9.02 | 16.54 | 26.23 | 10.47 |
| + Ours | 16.48 ▼ | 29.08 ▼ | 16.70 ▼ | 18.92 ▲ | 30.64 ▲ | 13.73 ▼ | 19.65 ▲ | 32.83 ▲ | 8.42 ▲ | 14.33 ▲ | 23.29 ▲ | 9.81 ▲ |
| STSGCN | 18.24 | 30.51 | 17.37 | 21.46 | 34.07 | 14.37 | 24.07 | 38.69 | 10.17 | 17.05 | 26.47 | 10.87 |
| + Ours | 17.40 ▲ | 30.71 ▼ | 17.97 ▼ | 19.34 ▲ | 31.36 ▲ | 13.28 ▲ | 20.35 ▲ | 33.97 ▲ | 8.62 ▲ | 15.26 ▲ | 24.49 ▲ | 10.04 ▲ |
| STAEformer | 16.67 | 29.56 | 17.21 | 18.63 | 30.86 | 13.10 | 19.91 | 33.94 | 8.58 | 13.82 | 23.55 | 9.23 |
| + Ours | 16.03 ▲ | 28.30 ▲ | 17.28 ▼ | 18.69 ▼ | 30.35 ▲ | 13.64 ▼ | 19.82 ▲ | 33.08 ▲ | 8.79 ▼ | 13.64 ▲ | 22.99 ▲ | 9.58 ▼ |
| DSTAGNN | 15.99 | 28.20 | 15.45 | 19.85 | 31.91 | 13.42 | 21.84 | 34.95 | 9.58 | 16.19 | 25.44 | 10.61 |
| + Ours | 16.44 ▼ | 28.39 ▼ | 18.39 ▼ | 19.53 ▲ | 31.31 ▲ | 14.58 ▼ | 21.14 ▲ | 34.00 ▲ | 9.68 ▼ | 14.89 ▲ | 23.58 ▲ | 10.73 ▼ |
| STWave | 15.91 | 28.06 | 16.62 | 19.63 | 32.02 | 13.66 | 21.32 | 35.68 | 8.99 | 14.01 | 23.94 | 9.33 |
| + Ours | 16.00 ▼ | 28.74 ▼ | 16.81 ▼ | 18.84 ▲ | 30.78 ▲ | 13.05 ▲ | 19.30 ▲ | 32.20 ▲ | 8.28 ▲ | 13.71 ▲ | 22.94 ▲ | 9.15 ▲ |
| STIDGCN | 15.98 | 26.66 | 17.24 | 18.38 | 30.47 | 12.42 | 19.20 | 32.87 | 8.18 | 13.71 | 23.33 | 9.00 |
| + Ours | 15.20 ▲ | 25.60 ▲ | 15.70 ▲ | 18.27 ▲ | 30.23 ▲ | 13.02 ▼ | 18.83 ▲ | 32.20 ▲ | 7.98 ▲ | 13.12 ▲ | 22.55 ▲ | 8.74 ▲ |
| DTRformer | 15.83 | 27.61 | 16.31 | 18.49 | 30.62 | 13.08 | 19.52 | 32.53 | 8.38 | 13.50 | 22.90 | 9.05 |
| + Ours | 17.43 ▼ | 31.00 ▼ | 17.85 ▼ | 19.46 ▼ | 31.65 ▲ | 13.70 ▼ | 19.92 ▼ | 32.91 ▼ | 8.72 ▼ | 14.24 ▼ | 23.40 ▼ | 9.61 ▼ |
| LHMformer | 16.22 | 27.81 | 16.76 | 18.45 | 30.70 | 13.00 | 19.27 | 32.54 | 8.11 | 13.66 | 23.04 | 9.02 |
| + Ours | 16.15 ▲ | 28.70 ▼ | 16.55 ▲ | 18.83 ▼ | 30.52 ▲ | 13.41 ▼ | 19.76 ▼ | 32.94 ▲ | 8.92 ▼ | 13.53 ▲ | 22.63 ▲ | 9.17 ▼ |
| SIMBAD | 14.77 | 25.87 | 14.35 | 18.03 | 30.08 | 12.36 | 17.86 | 30.99 | 7.44 | 12.58 | 22.13 | 8.35 |
Table 5. Performance comparison of baseline models with and without the proposed Aperiodicity-Aware Modulation Layer on PEMS datasets. Superscript on +Ours indicate improvement (▲) or degradation (▼).
Stretched Training Dataset
As in Section 2, conventional models are given previous observation of timesteps to predict future traffic flow of timesteps with input by learning a mapping function
where the previous observation of timesteps are referred to as recent input in our model. However, our model incorporates extra historical data (weekly and daily) as inputs for prediction. As a result, our model takes in more inputs than existing baselines to predict the same timesteps of traffic flow. This may result in an imbalance of training datasets between our model and existing ones which may lead to unfair comparison. To overcome this, we stretched the length of training datasets for baselines to match the scope of training data. Table 6 below shows results between our model and baselines with stretched training data.
| Model | PEMS03 | PEMS04 | PEMS07 | PEMS08 | ||||||||
| Metric | MAE | RMSE | MAPE% | MAE | RMSE | MAPE% | MAE | RMSE | MAPE% | MAE | RMSE | MAPE% |
| GWNET | 14.63 | 25.45 | 14.71 | 18.88 | 30.38 | 13.45 | 20.52 | 33.10 | 8.84 | 14.72 | 23.53 | 9.56 |
| DCRNN | 16.31 | 28.71 | 16.07 | 20.91 | 33.21 | 13.96 | 22.66 | 35.68 | 10.11 | 16.13 | 25.32 | 10.42 |
| AGCRN | 15.93 | 27.82 | 14.88 | 19.47 | 32.16 | 13.05 | 20.91 | 34.60 | 8.86 | 16.04 | 25.46 | 10.29 |
| STSGCN | 17.65 | 29.39 | 16.73 | 20.97 | 33.61 | 13.88 | 23.78 | 38.72 | 10.01 | 16.64 | 25.98 | 10.68 |
| STAEformer | 15.81 | 27.93 | 16.12 | 18.41 | 30.87 | 12.83 | 19.73 | 33.53 | 8.72 | 13.70 | 23.19 | 9.33 |
| DSTAGNN | 15.64 | 27.58 | 14.78 | 19.41 | 31.60 | 12.95 | 21.62 | 34.72 | 9.45 | 15.81 | 24.90 | 10.08 |
| STWave | 15.03 | 27.14 | 15.21 | 18.75 | 30.83 | 12.97 | 20.53 | 34.33 | 8.76 | 13.45 | 23.39 | 8.90 |
| STIDGCN | 15.50 | 25.91 | 17.05 | 18.39 | 30.52 | 12.69 | 19.02 | 32.74 | 7.99 | 13.50 | 23.21 | 9.00 |
| DTRformer | 15.27 | 27.14 | 15.64 | 18.21 | 30.11 | 12.75 | 19.32 | 32.31 | 8.25 | 13.35 | 22.80 | 8.90 |
| LHMformer | 15.49 | 25.88 | 15.82 | 18.17 | 30.62 | 12.75 | 19.14 | 32.44 | 8.04 | 13.32 | 22.78 | 8.78 |
| ASTGCN | 20.25 | 35.99 | 20.67 | 22.20 | 35.24 | 15.66 | 25.50 | 39.75 | 11.43 | 17.07 | 26.49 | 11.80 |
| SIMBAD | 14.77 | 25.87 | 14.35 | 18.03 | 30.08 | 12.36 | 17.86 | 30.99 | 7.44 | 12.58 | 22.13 | 8.35 |
Table 6. Performance comparison on the stretched training dataset.
Standard Deviation Results
In this section, we show standard deviation results of the overall performance results. We also compare with NLinear and DLinear [38], which have been shown to be competitive on traffic data with minimal parameters.
| Model | PEMS03 | PEMS04 | PEMS07 | PEMS08 | ||||||||
| Metric | MAE | RMSE | MAPE% | MAE | RMSE | MAPE% | MAE | RMSE | MAPE% | MAE | RMSE | MAPE% |
| NLinear | 21.07 ± 0.00 | 34.07 ± 0.01 | 19.50 ± 0.02 | 28.80 ± 0.00 | 45.47 ± 0.01 | 19.00 ± 0.01 | 31.28 ± 0.00 | 49.12 ± 0.01 | 13.17 ± 0.01 | 22.76 ± 0.00 | 36.12 ± 0.01 | 13.57 ± 0.01 |
| DLinear | 20.72 ± 0.01 | 33.28 ± 0.02 | 20.73 ± 0.14 | 28.34 ± 0.00 | 44.44 ± 0.01 | 20.63 ± 0.07 | 30.66 ± 0.03 | 47.77 ± 0.05 | 13.31 ± 0.05 | 22.44 ± 0.01 | 35.31 ± 0.02 | 13.78 ± 0.11 |
| GWNET | 14.94 ± 0.12 | 26.06 ± 0.38 | 15.52 ± 0.79 | 19.23 ± 0.09 | 30.76 ± 0.13 | 14.04 ± 0.51 | 20.63 ± 0.05 | 33.26 ± 0.07 | 9.14 ± 0.17 | 15.06 ± 0.11 | 23.79 ± 0.12 | 9.73 ± 0.02 |
| DCRNN | 16.67 ± 0.12 | 29.06 ± 0.33 | 16.60 ± 0.93 | 21.68 ± 0.18 | 34.19 ± 0.22 | 15.05 ± 0.51 | 23.13 ± 0.13 | 36.13 ± 0.14 | 10.18 ± 0.58 | 16.81 ± 0.07 | 25.91 ± 0.08 | 11.71 ± 0.58 |
| AGCRN | 16.21 ± 0.12 | 28.57 ± 0.39 | 15.35 ± 0.38 | 19.85 ± 0.18 | 32.96 ± 0.22 | 13.38 ± 0.14 | 21.35 ± 0.15 | 35.12 ± 0.14 | 9.02 ± 0.07 | 16.54 ± 0.07 | 26.23 ± 0.36 | 10.47 ± 0.19 |
| STSGCN | 18.24 ± 0.11 | 30.51 ± 0.52 | 17.37 ± 0.17 | 21.46 ± 0.06 | 34.07 ± 0.14 | 14.37 ± 0.07 | 24.07 ± 0.34 | 38.69 ± 0.07 | 10.17 ± 0.25 | 17.05 ± 0.17 | 26.47 ± 0.25 | 10.87 ± 0.11 |
| STAEformer | 16.67 ± 0.22 | 29.56 ± 0.42 | 17.21 ± 0.18 | 18.63 ± 0.03 | 30.86 ± 0.25 | 13.60 ± 0.17 | 19.91 ± 0.49 | 33.94 ± 0.08 | 8.58 ± 0.14 | 13.82 ± 0.09 | 23.55 ± 0.23 | 9.23 ± 0.14 |
| DSTAGNN | 15.99 ± 0.11 | 28.20 ± 0.15 | 15.45 ± 0.71 | 19.85 ± 0.19 | 31.91 ± 0.11 | 13.42 ± 0.30 | 21.84 ± 0.36 | 34.95 ± 0.32 | 9.58 ± 0.18 | 16.19 ± 0.25 | 25.44 ± 0.27 | 10.61 ± 0.27 |
| STWave | 15.91 ± 0.34 | 28.06 ± 0.16 | 16.62 ± 0.34 | 19.63 ± 0.51 | 32.02 ± 0.74 | 13.66 ± 0.38 | 21.32 ± 0.75 | 35.68 ± 0.13 | 8.99 ± 0.42 | 14.01 ± 0.39 | 23.94 ± 0.71 | 9.33 ± 0.42 |
| STIDGCN | 15.98 ± 0.11 | 26.66 ± 0.26 | 17.24 ± 0.34 | 18.38 ± 0.06 | 30.47 ± 0.18 | 12.42 ± 0.21 | 19.20 ± 0.11 | 32.87 ± 0.23 | 8.18 ± 0.08 | 13.71 ± 0.10 | 23.33 ± 0.03 | 9.00 ± 0.08 |
| DTRformer | 15.83 ± 0.16 | 27.61 ± 0.42 | 16.31 ± 0.28 | 18.49 ± 0.10 | 30.62 ± 0.29 | 13.08 ± 0.18 | 19.52 ± 0.14 | 32.53 ± 0.14 | 8.38 ± 0.14 | 13.50 ± 0.04 | 22.90 ± 0.09 | 9.05 ± 0.17 |
| LHMformer | 16.22 ± 0.09 | 27.81 ± 0.25 | 16.76 ± 0.28 | 18.45 ± 0.03 | 30.70 ± 0.08 | 13.00 ± 0.05 | 19.27 ± 0.05 | 32.54 ± 0.05 | 8.11 ± 0.01 | 13.66 ± 0.07 | 23.04 ± 0.02 | 9.02 ± 0.06 |
| ASTGCN | 20.25 ± 0.27 | 35.99 ± 0.53 | 20.67 ± 0.55 | 22.20 ± 0.26 | 35.24 ± 0.49 | 15.66 ± 0.59 | 25.50 ± 0.37 | 39.75 ± 0.66 | 11.43 ± 0.15 | 17.07 ± 0.21 | 26.49 ± 0.32 | 11.80 ± 0.34 |
| SIMBAD | 14.77 ± 0.04 | 25.87 ± 0.15 | 14.35 ± 0.13 | 18.03 ± 0.04 | 30.08 ± 0.06 | 12.36 ± 0.02 | 17.86 ± 0.02 | 30.99 ± 0.02 | 7.44 ± 0.02 | 12.58 ± 0.02 | 22.13 ± 0.03 | 8.35 ± 0.07 |
Table 7. Standard deviation results on PEMS datasets.
Hyperparameter Analysis
In this section, we compare two hyperparameter settings. (1) : a temperature parameter which controls the sharpness of distribution of Eq. (8). (2) : regularizes the loss of Eq. (26). Table 8 shows results in terms of MAE, RMSE, and MAPE on all four datasets. Our choice of is 0.2 and is 0.1.
| PEMS03 | PEMS04 | |||||
| MAE | RMSE | MAPE | MAE | RMSE | MAPE | |
| Default | 14.77 | 25.87 | 14.35 | 18.03 | 30.08 | 12.36 |
| temperature () | ||||||
| 0.1 | 14.99 | 26.27 | 14.84 | 18.04 | 30.11 | 12.38 |
| 0.3 | 14.99 | 26.30 | 14.66 | 18.03 | 30.08 | 12.37 |
| 0.4 | 15.03 | 26.35 | 14.96 | 18.04 | 30.08 | 12.35 |
| regularizer () | ||||||
| 0.2 | 14.92 | 26.30 | 14.62 | 18.04 | 30.09 | 12.40 |
| 0.3 | 14.98 | 26.38 | 14.79 | 18.02 | 30.06 | 12.32 |
| 0.4 | 14.90 | 26.28 | 14.61 | 18.06 | 30.09 | 12.36 |
| PEMS07 | PEMS08 | |||||
| MAE | RMSE | MAPE | MAE | RMSE | MAPE | |
| Default | 17.86 | 30.99 | 7.44 | 12.58 | 22.13 | 8.35 |
| temperature () | ||||||
| 0.1 | 17.86 | 30.99 | 7.44 | 12.60 | 22.15 | 8.38 |
| 0.3 | 17.85 | 30.98 | 7.43 | 12.60 | 22.10 | 8.40 |
| 0.4 | 17.86 | 30.98 | 7.44 | 12.62 | 22.12 | 8.43 |
| regularizer () | ||||||
| 0.2 | 17.86 | 30.99 | 7.45 | 12.61 | 22.15 | 8.43 |
| 0.3 | 17.88 | 31.00 | 7.47 | 12.63 | 22.15 | 8.46 |
| 0.4 | 17.90 | 31.02 | 7.48 | 12.67 | 22.15 | 8.45 |
Table 8. Performance of hyperparameter-tuned variants. Best results are marked in bold.
Computational Efficiency
In this section, we compare the computational complexity between baselines on the tested datasets. Compared to most recent baselines [5, 26], SIMBAD achieves more efficient training and inference while maintaining superior performance. This demonstrates that utilization of periodic patterns and uncertainty can improve forecasting accuracy without introducing excessive computational overhead.
| Model | PEMS03 | PEMS04 | PEMS07 | PEMS08 | ||||
| Metric | Train (s) | Test (s) | Train (s) | Test (s) | Train (s) | Test (s) | Train (s) | Test (s) |
| GWNET | 8.33 | 1.43 | 4.17 | 0.91 | 27.12 | 4.00 | 3.03 | 0.70 |
| DCRNN | 51.69 | 10.80 | 28.03 | 5.52 | 105.07 | 44.47 | 25.76 | 4.77 |
| AGCRN | 17.82 | 2.93 | 10.30 | 1.85 | 27.45 | 3.96 | 11.10 | 1.91 |
| STSGCN | 23.68 | 6.56 | 11.58 | 2.87 | 97.44 | 21.29 | 12.37 | 1.56 |
| STAEformer | 40.46 | 4.33 | 19.78 | 2.29 | 137.60 | 14.28 | 11.26 | 1.58 |
| DSTAGNN | 65.79 | 7.99 | 30.09 | 3.69 | 280.68 | 30.01 | 14.42 | 1.91 |
| STWave | 47.70 | 7.38 | 24.05 | 4.15 | 141.10 | 21.13 | 14.80 | 2.72 |
| STIDGCN | 23.65 | 3.15 | 25.74 | 1.94 | 187.64 | 19.98 | 9.08 | 1.19 |
| DTRformer | 61.97 | 6.28 | 30.54 | 3.11 | 267.35 | 27.41 | 18.14 | 2.14 |
| LHMformer | 421.42 | 8.08 | 45.24 | 3.28 | 228.11 | 27.54 | 18.98 | 2.25 |
| ASTGCN | 31.11 | 3.01 | 18.51 | 1.55 | 68.63 | 10.39 | 19.16 | 1.18 |
| SIMBAD | 36.36 | 4.97 | 17.15 | 2.58 | 140.84 | 19.00 | 9.55 | 1.74 |
Table 9. Train per epoch and testing time comparison.
Additional Datasets
We compare the performance of our model with the baselines on additional datasets METR-LA, PEMS-BAY, and a subset of LargeST [24] datasets. For the less efficient baselines, we use one year of traffic data from 2019. The details of the datasets are given in Table 10 and the results are given in Table 11. We split the METR-LA and PEMS-BAY datasets into train, validation and test sets in a 7 : 1 : 2 ratio and SD dataset into 6 : 2 : 2 ratio.
| DATASET | NODES | EDGES | TIMESTEPS |
| METR-LA | 207 | 1,515 | 34,272 |
| PEMS-BAY | 325 | 2,369 | 52,116 |
| SD (2019) | 716 | 17,319 | 35,040 |
Table 10. Summary of datasets.
| DATASET | METRIC | GWNET | DCRNN | AGCRN | STSGCN | STAEformer | DSTAGNN | STWave | STIDGCN | DTRformer | LHMformer | ASTGCN | SIMBAD |
| SD (2019) Average | MAE | 19.90 | 21.74 | 23.04 | 204.60 | 18.86 | 21.74 | 19.75 | 18.47 | 19.71 | 19.59 | 23.60 | 17.98 |
| RMSE | 32.63 | 35.63 | 40.69 | 86.63 | 32.99 | 38.28 | 32.35 | 31.30 | 34.30 | 33.33 | 41.48 | 30.15 | |
| MAPE | 13.54 | 14.61 | 15.17 | 268.05 | 12.48 | 13.56 | 13.25 | 12.18 | 12.93 | 12.81 | 16.08 | 11.79 | |
| PEMS-BAY Horizon 3 (15 min) | MAE | 1.31 | 1.35 | 1.40 | 1.51 | 1.36 | 1.42 | 1.37 | 1.34 | 1.38 | 1.34 | 1.82 | 1.28 |
| RMSE | 2.77 | 2.86 | 2.98 | 3.27 | 2.90 | 3.05 | 2.98 | 2.88 | 2.88 | 2.83 | 3.55 | 2.76 | |
| MAPE | 2.74 | 2.85 | 3.00 | 3.77 | 2.90 | 3.12 | 2.93 | 2.86 | 2.96 | 2.85 | 4.09 | 2.72 | |
| PEMS-BAY Horizon 6 (30 min) | MAE | 1.66 | 1.70 | 1.75 | 1.87 | 1.67 | 1.77 | 1.71 | 1.67 | 1.70 | 1.66 | 2.10 | 1.58 |
| RMSE | 3.78 | 3.90 | 3.97 | 4.31 | 3.86 | 4.05 | 3.97 | 3.87 | 3.78 | 3.79 | 4.26 | 3.67 | |
| MAPE | 3.74 | 3.85 | 4.05 | 4.32 | 3.82 | 4.08 | 3.90 | 3.81 | 3.93 | 3.79 | 4.92 | 3.59 | |
| PEMS-BAY Horizon 12 (60 min) | MAE | 2.00 | 2.05 | 2.11 | 2.30 | 1.96 | 2.12 | 2.02 | 1.94 | 1.99 | 1.93 | 2.43 | 1.84 |
| RMSE | 4.62 | 4.77 | 4.81 | 5.50 | 4.56 | 4.89 | 4.72 | 4.55 | 4.46 | 4.44 | 4.90 | 4.32 | |
| MAPE | 4.73 | 4.88 | 5.06 | 5.85 | 4.69 | 5.04 | 4.84 | 4.60 | 4.80 | 4.57 | 5.76 | 4.40 | |
| METR-LA Horizon 3 (15 min) | MAE | 2.71 | 2.79 | 3.53 | 3.34 | 2.76 | 3.80 | 2.84 | 2.79 | 2.80 | 2.66 | 4.94 | 2.79 |
| RMSE | 5.18 | 5.37 | 9.25 | 8.22 | 5.41 | 9.38 | 5.65 | 5.54 | 5.49 | 5.16 | 9.80 | 5.49 | |
| MAPE | 6.95 | 7.25 | 8.35 | 7.58 | 7.23 | 8.15 | 7.56 | 7.52 | 7.45 | 6.86 | 9.87 | 7.33 | |
| METR-LA Horizon 6 (30 min) | MAE | 3.10 | 3.20 | 4.56 | 4.12 | 3.09 | 4.81 | 3.24 | 3.16 | 3.14 | 2.97 | 6.39 | 3.18 |
| RMSE | 6.23 | 6.50 | 11.70 | 10.39 | 6.39 | 12.22 | 6.76 | 6.60 | 6.46 | 6.06 | 12.05 | 6.55 | |
| MAPE | 8.43 | 8.75 | 10.60 | 9.63 | 8.57 | 10.17 | 9.25 | 9.05 | 8.86 | 8.11 | 12.05 | 8.88 | |
| METR-LA Horizon 12 (60 min) | MAE | 3.57 | 3.73 | 6.04 | 5.16 | 3.48 | 6.29 | 3.61 | 3.52 | 3.51 | 3.34 | 6.39 | 3.66 |
| RMSE | 7.39 | 7.78 | 14.70 | 13.38 | 7.05 | 15.23 | 7.70 | 7.51 | 7.39 | 7.05 | 14.76 | 7.57 | |
| MAPE | 10.10 | 10.73 | 13.50 | 11.90 | 10.14 | 12.73 | 10.86 | 10.54 | 10.34 | 9.63 | 14.44 | 10.66 |
Table 11. Performance comparison on SD (2019), PEMS-BAY, and METR-LA.
SIMBAD achieves improvement of 3.18% on average over the best baselines on the SD dataset and 2.91% on the PEMS-BAY dataset. However, compared to the superior performance on PEMS datasets, SIMBAD shows relatively low performance on METR-LA. In contrast to the clear positive correlation between L1-norm of current-historical and L1-norm of target-historical data of PEMS04 as illustrated in Figure 9(a), METR-LA shows a substantially weaker correlation between the two L1-norms, as in Figure 9(b). This suggests that the current state is less informative for selecting reliable historical patterns, violating a key assumption of similarity-based aggregation. As a result, the model’s ability to identify and exploit useful periodic information is diminished, leading to reduced performance gains.

Figure 9. Correlation between L1-norms of input and target sequences with respect to their last week counterparts. (a) PEMS04. (b) METR-LA.
Baseline Details
GWNET [35] combines gated temporal convolution layer (TCN) and adaptive graph, which is learned from node embeddings, based on GCN. Long-range temporal sequences are handled by stacked dilated 1D causal convolutions.
DCRNN [21] leverages diffusion process on a graph and capture spatial dependencies via bidirectional random walks. Temporal dependencies are handled by Gated Recurrent Units (GRUs), which are replaced by the diffusion convolutions to effectively capture spatio-temporal dependencies.
AGCRN [1] integrates graph convolution with recurrent modeling to capture spatio-temporal dependencies in traffic data.
STSGCN [30] synchronously captures localized spatio-temporal correlations by constructing a single localized spatio-temporal graph that links nodes across adjacent timesteps.
STAEformer [23] designed novel spatio-temporal adaptive embeddings which are fed into stacked temporal and spatial vanilla Transformer layers.
DSTAGNN [18] utilizes a data-driven graph constructed via Wasserstein distance and employs graph convolutions and multi-scale gated convolutions (M-GTU) to capture diverse spatial and temporal dynamics, respectively.
STWave [10] disentangled trend and event components with wavelet transform to capture both high- and low-frequency components
STIDGCN [22] incorporates a traffic pattern bank due to spatial heterogeneity and generates a data-driven dynamic graph.
DTRformer [5] utilizes adaptive embedding and cross spatial-temporal attention to fuse dynamic and static information.
LHMformer [26] combines short-term dependencies and long historical patterns through Transformer and TCN MLP module, respectively.
ASTGCN [13] incorporates weekly and daily data with attention mechanism to extract periodic patterns.
NeurIPS Paper Checklist
Claims
Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope?
Answer: [Yes]
Justification: The claims are consistent with the contributions and scope as detailed in the abstract and introduction.
Guidelines:
The answer [N/A] means that the abstract and introduction do not include the claims made in the paper.
The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A [No] or [N/A] answer to this question will not be perceived well by the reviewers.
The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings.
It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.
Limitations
Question: Does the paper discuss the limitations of the work performed by the authors?
Answer: [Yes]
Justification: The existing limitation of our work is given in the appendix.
Guidelines:
The answer [N/A] means that the paper has no limitation while the answer [No] means that the paper has limitations, but those are not discussed in the paper.
The authors are encouraged to create a separate “Limitations” section in their paper.
The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be.
The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated.
The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon.
The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size.
If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness.
While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren’t acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.
Theory assumptions and proofs
Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?
Answer: [N/A] Justification: The paper does not include theoretical claim.
Guidelines:
The answer [N/A] means that the paper does not include theoretical results.
All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced.
All assumptions should be clearly stated or referenced in the statement of any theorems.
The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition.
Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material.
Theorems and Lemmas that the proof relies upon should be properly referenced.
Experimental result reproducibility
Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?
Answer: [Yes]
Justification: The experimental setup and datasets are detailed in Section 4 and the proposed architecture is outlined in § 3. Train/test details and model hyperparameters are included in the main text.
Guidelines:
The answer [N/A] means that the paper does not include experiments.
If the paper includes experiments, a [No] answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not.
If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable.
Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general. releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed.
While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example (a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. (b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. (c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). (d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.
Open access to data and code
Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
Answer: [Yes]
Justification: Access to code is included in the introduction section.
Guidelines:
The answer [N/A] means that paper does not include experiments requiring code.
Please see the NeurIPS code and data submission guidelines (
https://neurips.cc/public/guides/CodeSubmissionPolicy) for more details.While we encourage the release of code and data, we understand that this might not be possible, so [No] is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark).
The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (
https://neurips.cc/public/guides/CodeSubmissionPolicy) for more details.The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc.
The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why.
At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable).
Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.
Experimental setting/details
Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer) necessary to understand the results?
Answer: [Yes]
Justification: Data description with train/val/test splits and hyperparameters are given in the main text.
Guidelines:
The answer [N/A] means that the paper does not include experiments.
The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them.
The full details can be provided either with the code, in appendix, or as supplemental material.
Experiment statistical significance
Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
Answer: [Yes]
Justification: The standard deviation error on the results is given in the appendix.
Guidelines:
The answer [N/A] means that the paper does not include experiments.
The authors should answer [Yes] if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.
The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions).
The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.)
The assumptions made should be given (e.g., Normally distributed errors).
It should be clear whether the error bar is the standard deviation or the standard error of the mean.
It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a 96% CI, if the hypothesis of Normality of errors is not verified.
For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g., negative error rates).
If error bars are reported in tables or plots, the authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.
Experiments compute resources
Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?
Answer: [Yes]
Justification: Time of inference and test is given in the appendix.
Guidelines:
The answer [N/A] means that the paper does not include experiments.
The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage.
The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute.
The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn’t make it into the paper).
Code of ethics
Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?
Answer: [Yes]
Justification: The research adheres to the NeurIPS Code of Ethics.
Guidelines:
The answer [N/A] means that the authors have not reviewed the NeurIPS Code of Ethics.
If the authors answer [No], they should explain the special circumstances that require a deviation from the Code of Ethics.
The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).
Broader impacts
Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
Answer: [Yes]
Justification: Our work can positively contribute to smarter and more reliable transportation systems, by improving robustness under irregular and dynamically changing traffic patterns.
Guidelines:
The answer [N/A] means that there is no societal impact of the work performed.
If the authors answer [N/A] or [No], they should explain why their work has no societal impact or why the paper does not address societal impact.
Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations.
The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate Deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster.
The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology.
If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).
Safeguards
Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pre-trained language models, image generators, or scraped datasets)?
Answer: [N/A]
Justification: The paper poses no such risk.
Guidelines:
The answer [N/A] means that the paper poses no such risks.
Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters.
Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images.
We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.
Licenses for existing assets
Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
Answer: [Yes]
Justification: The original creators have been attributed in the form of citations.
Guidelines:
The answer [N/A] means that the paper does not use existing assets.
The authors should cite the original paper that produced the code package or dataset.
The authors should state which version of the asset is used and, if possible, include a URL.
The name of the license (e.g., CC-BY 4.0) should be included for each asset.
For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.
If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset.
For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided.
If this information is not available online, the authors are encouraged to reach out to the asset’s creators.
New assets
Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
Answer: [N/A]
Justification: No new assets are included in the paper.
Guidelines:
The answer [N/A] means that the paper does not release new assets.
Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc.
The paper should discuss whether and how consent was obtained from people whose asset is used.
At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.
Crowdsourcing and research with human subjects
Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?
Answer: [N/A]
Justification: The research does not involve crowdsourcing experiments or research with human objects.
Guidelines:
The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects.
Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper.
According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.
Institutional review board (IRB) approvals or equivalent for research with human subjects
Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?
Answer: [N/A]
Justification: The paper does not involve research with human subjects.
Guidelines:
The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects.
Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.
We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution.
For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.
Declaration of LLM usage
Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigor, or originality of the research, declaration is not required.
Answer: [N/A]
Justification: The core method development in this work does not involve LLMs as any important, original or non-standard components.
Guidelines:
The answer [N/A] means that the core method development in this research does not involve LLMs as any important, original, or non-standard components.
Please refer to our LLM policy in the NeurIPS handbook for what should or should not be described.
References
- [1]Lei Bai, Lina Yao, Can Li, Xianzhi Wang, and Can Wang. Adaptive graph convolutional recurrent network for traffic forecasting. Advances in neural information processing systems, 33:17804–17815, 2020.arxiv.org/abs/2007.02842
- [2]Donald J Berndt and James Clifford. Using dynamic time warping to find patterns in time series. In Proceedings of the 3rd international conference on knowledge discovery and data mining, pages 359–370, 1994.
- [3]Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs. arXiv preprint arXiv:1312.6203, 2013.arxiv.org/abs/1312.6203
- [4]Ling Cai, Krzysztof Janowicz, Gengchen Mai, Bo Yan, and Rui Zhu. Traffic transformer: Capturing the continuity and periodicity of time series for traffic forecasting. Transactions in GIS, 24(3):736–755, 2020.DOI
- [5]Jing Chen, Haocheng Ye, Zhian Ying, Yuntao Sun, and Wenqiang Xu. Dynamic trend fusion module for traffic flow prediction. Applied Soft Computing, 174:112979, 2025.DOI
- [6]Yann N. Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated convolutional networks. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 933–941. PMLR, 06–11 Aug 2017.
- [7]Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. Advances in neural information processing systems, 29, 2016.arxiv.org/abs/1606.09375
- [8]Harris Drucker, Christopher J Burges, Linda Kaufman, Alex Smola, and Vladimir Vapnik. Support vector regression machines. Advances in neural information processing systems, 9, 1996.
- [9]Shengdong Du, Tianrui Li, Xun Gong, Yan Yang, and Shi Jinn Horng. Traffic flow forecasting based on hybrid deep learning framework. In 2017 12th international conference on intelligent systems and knowledge engineering (ISKE), pages 1–6. IEEE, 2017.DOI
- [10]Yuchen Fang, Yanjun Qin, Haiyong Luo, Fang Zhao, Bingbing Xu, Liang Zeng, and Chenxing Wang. When spatio-temporal meet wavelets: Disentangled traffic forecasting via efficient spectral graph attention networks. In 2023 IEEE 39th International Conference on Data Engineering (ICDE), pages 517–529. IEEE, 2023.DOI
- [11]Yuchen Fang, Yuxuan Liang, Bo Hui, Zhezhi Shao, Liwei Deng, Xu Liu, Xinke Jiang, and Kai Zheng. Efficient large-scale traffic forecasting with transformers: A spatial data management perspective. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1, pages 307–317, 2025.arxiv.org/abs/2412.09972
- [12]Marta C Gonzalez, Cesar A Hidalgo, and Albert-Laszlo Barabasi. Understanding individual human mobility patterns. nature, 453(7196):779–782, 2008.
- [13]Shengnan Guo, Youfang Lin, Ning Feng, Chao Song, and Huaiyu Wan. Attention based spatial-temporal graph convolutional networks for traffic flow forecasting. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 922–929, 2019.DOI
- [14]Jiawei Jiang, Chengkai Han, Wayne Xin Zhao, and Jingyuan Wang. Pdformer: Propagation delay-aware dynamic long-range transformer for traffic flow prediction. In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 4365–4373, 2023.arxiv.org/abs/2301.07945
- [15]Renhe Jiang, Zhaonan Wang, Jiawei Yong, Puneet Jеph, Quanjun Chen, Yasumasa Kobayashi, Xuan Song, Shintaro Fukushima, and Toyotaro Suzumura. Spatio-temporal meta-graph learning for traffic forecasting. In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 8078–8086, 2023.arxiv.org/abs/2211.14701
- [16]Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017.arxiv.org/abs/1703.04977
- [17]Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016.arxiv.org/abs/1609.02907
- [18]Shiyong Lan, Yitong Ma, Weikang Huang, Wenwu Wang, Hongyu Yang, and Pyang Li. Dstagnn: Dynamic spatial-temporal aware graph neural network for traffic flow forecasting. In International conference on machine learning, pages 11906–11917. PMLR, 2022.
- [19]Hyunwook Lee, Seungmin Jin, Hyeshin Chu, Hongkyu Lim, and Sungahn Ko. Learning to remember patterns: Pattern matching memory networks for traffic forecasting. arXiv preprint arXiv:2110.10380, 2021.arxiv.org/abs/2110.10380
- [20]Mengzhang Li and Zhanxing Zhu. Spatial-temporal fusion graph neural networks for traffic flow forecasting. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 4189–4196, 2021.
- [21]Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926, 2017.arxiv.org/abs/1707.01926
- [22]Aoyu Liu and Yaying Zhang. Spatial–temporal dynamic graph convolutional network with interactive learning for traffic forecasting. IEEE Transactions on Intelligent Transportation Systems, 25(7):7645–7660, 2024.DOI
- [23]Hangchen Liu, Zheng Dong, Renhe Jiang, Jiewen Deng, Jinliang Deng, Quanjun Chen, and Xuan Song. Spatio-temporal adaptive embedding makes vanilla transformer sota for traffic forecasting. In Proceedings of the 32nd ACM international conference on information and knowledge management, pages 4125–4129, 2023.
- [24]Xu Liu, Yutong Xia, Yuxuan Liang, Junfeng Hu, Yiwei Wang, Lei Bai, Chao Huang, Zhenguang Liu, Bryan Hooi, and Roger Zimmerman. Largest: A benchmark dataset for large-scale traffic forecasting. Advances in Neural Information Processing Systems, 36:75354–75371, 2023.arxiv.org/abs/2306.08259
- [25]Clélia Lopez, Ludovic Leclercq, Panchamy Krishnakumari, Nicolas Chiabaut, and Hans Van Lint. Revealing the day-to-day regularity of urban congestion patterns with 3d speed maps. Scientific Reports, 7(1):14029, 2017.DOI
- [26]Guoping Qi and Jihong Liu. Lhmformer: Long-range historical memory-enhanced transformer for traffic forecasting. In Proceedings of the 34th ACM International Conference on Information and Knowledge Management, pages 5125–5129, 2025.
- [27]Zezhi Shao, Zhao Zhang, Fei Wang, Wei Wei, and Yongjun Xu. Spatial-temporal identity: A simple yet effective baseline for multivariate time series forecasting. In Proceedings of the 31st ACM international conference on information & knowledge management, pages 4454–4458, 2022.arxiv.org/abs/2208.05233
- [28]David I Shuman, Sunil K Narang, Pascal Frossard, Antonio Ortega, and Pierre Vandergheynst. The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains. IEEE signal processing magazine, 30(3):83–98, 2013.DOI
- [29]Archana Singh, Avantika Yadav, and Ajay Rana. K-means with three different distance metrics. International Journal of Computer Applications, 67(10), 2013.DOI
- [30]Chao Song, Youfang Lin, Shengnan Guo, and Huaiyu Wan. Spatial-temporal synchronous graph convolutional networks: A new framework for spatial-temporal network data forecasting. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 914–921, 2020.DOI
- [31]Chaoming Song, Zehui Qu, Nicholas Blumm, and Albert-László Barabási. Limits of predictability in human mobility. Science, 327(5968):1018–1021, 2010.
- [32]Le Sun, Wenzhang Dai, and Ghulam Muhammad. Multi-level graph memory network cluster convolutional recurrent network for traffic forecasting. Information Fusion, 105:102214, 2024.
- [33]Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. arXiv preprint arXiv:1410.3916, 2014.arxiv.org/abs/1410.3916
- [34]Billy M Williams and Lester A Hoel. Modeling and forecasting vehicular traffic flow as a seasonal arima process: Theoretical basis and empirical results. Journal of transportation engineering, 129(6):664–672, 2003.DOI
- [35]Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and Chengqi Zhang. Graph wavenet for deep spatial-temporal graph modeling. arXiv preprint arXiv:1906.00121, 2019.arxiv.org/abs/1906.00121
- [36]ChuanTao Yin, Zhang Xiong, Hui Chen, Jing Yuan Wang, Daven Cooper, and Bertrand David. A literature survey on smart cities. Science China Information Sciences, 58(10):1–18, 2015.DOI
- [37]Bing Yu, Haoteng Yin, and Zhanxing Zhu. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. arXiv preprint arXiv:1709.04875, 2017.DOI
- [38]Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series forecasting? In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 11121–11128, 2023.arxiv.org/abs/2205.13504
- [39]Junbo Zhang, Yu Zheng, and Dekang Qi. Deep spatio-temporal residual networks for citywide crowd flows prediction. In Proceedings of the AAAI conference on artificial intelligence, volume 31, 2017.arxiv.org/abs/1610.00081
- [40]Chuanpan Zheng, Xiaoliang Fan, Cheng Wang, and Jianzhong Qi. Gman: A graph multi-attention network for traffic prediction. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 1234–1241, 2020.arxiv.org/abs/1911.08415
- [41]Eric Zivot and Jiahui Wang. Vector autoregressive models for multivariate time series. In Modeling financial time series with S-PLUS®, pages 369–413. Springer, 2006.