DRAN: A Distribution and Relation Adaptive Network for Spatio-temporal Forecasting

arXiv:2504.01531 · cs.LG · Submitted 2025-04-02 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "DRAN: A Distribution and Relation Adaptive Network for Spatio-temporal Forecasting".

Jane: The DRAN framework addresses challenges in spatio-temporal forecasting by integrating a Spatial Factor Learner (SFL) for distribution adaptation and a Dynamic-Static Fusion Learner (DSFL) to model evolving spatial dependencies.

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So we've talked about what DRAN is trying to achieve, which is handling those evolving spatial dependencies better than current methods. The paper starts by saying that spatio-temporal forecasting stays difficult because both the data distributions and spatial relations are constantly shifting over time >

Jane: And they point out that using temporal normalization and de-normalization to fix distribution shifts can end up distorting those inter-node relationships, which hurts the ability to model spatial dependencies properly >

Lu: To solve this, they introduce DRAN, which has two key modules: the SFL module for distribution adaptation and the DSFL module for modeling evolving spatial interactions >

Meng: The SFL module's job is essentially to normalize and de-normalize features in a way that respects the original spatial structures of the system >

Tom: It achieves this by learning spatial scaling and shifting factors, which means it generates these spatially consistent features, Hspa, through a specific de-normalization process >

Jane: And then they pair that up with the DSFL module which is designed to decompose the features into static and dynamic components >

Lu: The DSFL module uses a GCN layer called FGCN1 which acts like a low-pass filter in the graph spectral domain to get those static components, zSt, and then they use attention mechanisms for the dynamic component, zDy >

Tom: It seems like that decomposition lets them model spatial dependencies more thoroughly because they're separating what’s persistent from what’s changing right now >

Jane: And the final step is taking those static and dynamic features and fusing them using a gating mechanism conditioned on both the static features and metrics reflecting the input variability >

Meng: That gating mechanism is interesting because it means the model actively decides whether to lean more on the static spatial context or switch focus to the dynamic variations depending on what's happening in that specific prediction window >

Lu: It’s about regulating those relative contributions of static and dynamic representations according to current spatio-temporal variability, allowing the model to emphasize dynamic relations during fast changes while keeping stable structures under steady conditions >

Tom: So, DRAN is essentially a framework that integrates SFL for distribution adaptation and DSFL for adaptive modeling of static and dynamic spatial dependencies >

Jane: It claims that experiments on six benchmark datasets show DRAN outperforms state-of-the-art baselines across the board, which means it performs better on both the point error metrics and those distributional evaluation metrics too >

Conclusion: Tom: So we've covered how DRAN works by looking at its two main components, the SFL and DSFL modules that work together to handle distribution shifts and dynamic spatial interactions >

Jane: The authors of this paper, Xiaobei Zou, Luolin Xiong, Kexuan Zhang, Cesare Alippi, Fellow IEEE, Yang Tang Fellow IEEE—they put forward this framework because they felt existing methods weren't sophisticated enough for these non-stationary environments >

Lu: What the paper really implies is that featurewise normalization alone isn't sufficient for distribution adaptation in spatio-temporal tasks >

Meng: They found that applying SFL after temporal normalization provides a more effective way to adapt the distribution compared to other methods like instance or spatial level normalization, which suggests a different way to approach these shifts >

Tom: It’s about finding the right way to align those input distributions with the target distributions while preserving what's important for forecasting >

Jane: The overall message is that this combination of SFL and DSFL gives us a more comprehensive representation learning pathway for spatio-temporal systems than what we had before >

Lu: It shows that decomposing spatial information into static and dynamic parts and fusing them adaptively based on variability is a powerful concept for handling real-world, changing data streams >

Tom: We saw the results on those six datasets—Weather, NYCBike1, NYCBike2, NYCTaxi, PeMS04, and PeMS08—and DRAN consistently delivered the best overall results when compared to the baselines >

Jane: So for anyone who only listens to the show today, just know that DRAN is a framework that uses spatial factor learning and dynamic-static fusion to tackle those tricky spatio-temporal forecasting problems >

East China University of Science and Technology

cs.LG

Submitted: 2025-04-02

Updated: 2026-09-28

Importance score: 79/100

The gist: The DRAN framework addresses challenges in spatio-temporal forecasting by integrating a Spatial Factor Learner (SFL) for distribution adaptation and a Dynamic-Static Fusion Learner (DSFL) to model

Key concepts

Spatial Factor Learner (SFL)
This module learns spatial scaling and shifting factors to restore the original spatial structures that get distorted during temporal normalization. It ensures that the learned features maintain consistent spatial relationships without interfering with other dependency modeling, effectively aligning the data distribution.
Dynamic-Static Fusion Learner (DSFL)
DSFL decomposes input features into static components (persistent spatial structures) and dynamic components (sample-specific variations). It adaptively fuses these two parts based on current spatio-temporal variability, allowing the model to emphasize dynamic relations when things change rapidly.
Adaptive Fusion Mechanism
The final representation is created by using a gating mechanism conditioned on static features and input-dependent metrics. This mechanism decides how much influence the static versus dynamic spatial information should have at any given time, optimizing feature combination for the specific input conditions.

Terminology

Summary

The DRAN framework addresses challenges in spatio-temporal forecasting by integrating a Spatial Factor Learner (SFL) for distribution adaptation and a Dynamic-Static Fusion Learner (DSFL) to model evolving spatial dependencies. The gist is that DRAN outperforms state-of-the-art baselines on six benchmark datasets through its novel approach to distribution adaptation and adaptive fusion of static and dynamic spatial components.

How it works

DRAN incorporates a Spatial Factor Learner (SFL) module which enables effective normalization and de-normalization while preserving spatial dependencies in spatio-temporal systems. This module learns spatial scaling and shifting factors to restore the spatial structures affected by temporal normalization, explicitly considering spatial relational consistency without interfering with subsequent spatial dependency modeling. The SFL module generates the spatially consistent features Hspa by applying a de-normalization process where Hspa = σspaHtem + µspa. A reconstruction constraint is imposed to align Hspa with the raw features, encouraging SFL to preserve the spatial relations distorted by temporal normalization.

How it works

DRAN further proposes the Dynamic-Static Fusion Learner (DSFL) module to model evolving spatial interactions by decomposing features into static and dynamic components and adaptively fusing them according to input variability. This contrasts with existing methods that typically combine different components through summation, concatenation, or globally shared fusion weights. DSFL regulates the relative contributions of static and dynamic representations according to current spatio-temporal variability, allowing the model to emphasize dynamic relations during rapidly varying periods while retaining persistent spatial structures under relatively stable conditions. The decomposition involves a GCN layer FGCN1 which acts as a low-pass filter in the graph spectral domain to derive the static component zSt and the dynamic component zDy.

How it works

The DSFL module models spatial dependencies through attention mechanisms together with an adaptive adjacency matrix ASt constructed from learnable node embeddings. The static component zSt is obtained through graph low-pass filtering and is expected to capture persistent spatial structures shared across samples. The dynamic component zDy is explicitly defined as the residual part after removing zSt from Hspa, representing high-frequency or sample-specific spatial variations.

How it works

The final spatio-temporal representation H is obtained by adaptively fusing the static and dynamic features using a gating mechanism conditioned on static features and input-dependent dynamic variability metrics. The control feature zcontrol is generated by applying a Sigmoid activation function to the concatenation of the static features HSt, variance ratio of first differences zdvr, and high-frequency energy ratio zhf. The final fusion is achieved via H = zcontrol ⊙HDy + (1−zcontrol)⊙HSt.

How it works

After obtaining the encoded spatio-temporal representation H, a VAE is utilized to generate the stochastic components of the predictions by mapping H into latent layers to obtain the mean µsto and standard deviation σsto. These latent features zl are then sampled from a distribution N (µsto,σ squared sto) and processed through a reconstruction layer Frec to map back to the stochastic components of the time series HS. The overall loss function L consists of three components: prediction error Lpred, spatial similarity loss Lspa, and the distribution loss Lkl.

How it works

The model is trained using the Adam optimizer with a learning rate of 0.001, batch size 32, and 100 training epochs. The hyperparameters α and β in Eq. (14) are fine-tuned experimentally to account for the stochastic nature and uncertainties of the datasets.

How it works

The experiments evaluate DRAN on six real-world datasets covering weather and traffic systems: Weather, NYCBike1, NYCBike2, NYCTaxi, PeMS04, and PeMS08. Model performance is evaluated using Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE), and a distributional metric based on the Wasserstein Distance (WD). DRAN consistently outperforms the baselines, achieving the best overall results on both point error metrics and distributional evaluation metrics.

How it works

The ablation studies show that removing SFL leads to performance degradation across all datasets, particularly on the WD metric, indicating that SFL plays an important role in distribution alignment and also benefits point prediction accuracy. Removing the entire DSFL module results in a substantial decrease in prediction accuracy. The ablation results show that the model incorporating all modules achieves the best overall performance.

How it works

DRAN shows two limitations: Scalability, where online inference complexity is dominated by O((k+2)LN2C′), and Adaptation to abrupt spatial changes, as the static branch is learned from training data and the dynamic branch relies on a finite lookback window. Future work will explore sparse or low-rank spatial operators to reduce this overhead. The paper concludes that DRAN integrates a Spatial Factor Learner (SFL) for distribution adaptation and a Dynamic-Static Fusion Learner (DSFL) to model both dynamic and static spatial dependencies.

How it works

The paper demonstrates that SFL reduces PDD and WD on all datasets, showing that it generally mitigates the spatial distortion introduced by temporal normalization. SFL also improves the performance of various temporal normalization methods, demonstrating its generality in preserving spatial distributions. The model incorporating temporal operations and SFL outperforms the model with RevIN. This finding suggests that the featurewise normalization alone is too coarse-grained and is insufficient for distribution adaptation in spatio-temporal tasks. Furthermore, applying SFL after temporal normalization provides an effective approach for distribution adaptation compared with instance-level, spatial-level and temporal-level normalization.

Improvements for AI systems

  1. textbf Spatial Factor Learner for Distribution Adaptation (SFL): Mitigate Spatial Distortion During Normalization. The SFL module learn[s] spatial scaling and shifting factors to restore the spatial structures affected by temporal normalization, explicitly addressing how temporal normalization rescales nodes to the same mean and variance level (grey), causing the learned spatial interactions to deviate from the original spatial relationships.

  2. textbf Dynamic-Static Fusion Learner for Adaptive Spatial Modeling (DSFL): Capture Complementary Spatial Dependencies. DSFL decomposes features into static and dynamic components and adaptively fuses them according to input variability, allowing the model to emphasize dynamic relations during rapidly varying periods while retaining persistent spatial structures under relatively stable conditions.

  3. textbf Enhanced Spatio-Temporal Forecasting Accuracy: Achieve Superior Error Metrics. DRAN consistently outperforms baselines, achieving the best overall results on both point error metrics and distributional evaluation metrics, indicating it not only minimizes absolute prediction errors but also improves the distributional alignment between predictions and ground-truth values.

  4. textbf Robustness to Test-Time Perturbations: Maintain Performance Under Corruption. DRAN consistently achieves lower MAE than the strongest competing baseline across all evaluated conditions when tested under Gaussian noise and missing sensors, demonstrating that it maintains a consistent relative advantage under severe test-time corruptions.

  5. textbf Improved Spatial Relation Preservation: Minimize Pairwise Distance Distortion. The constraint loss Lspa is designed to encourage SFL to preserve the spatial relations distorted by temporal normalization, as shown by the theoretical result that the pairwise distance distortion is bounded by a term related to the temporal normalization discrepancy.

Sources

Related papers