ST-Hyper: Learning High-Order Dependencies Across Multiple Spatial-Temporal Scales for Multivariate Time Series Forecasting

Binqing Wu, Jianlong Huang, Zongjiang Shang, Ling Chen

Abstract

Multivariate time series (MTS) forecasting is a critical task in domains such as urban planning, energy management, and environmental monitoring. However, capturing the complex, high-order dependencies that exist across varying spatial and temporal scales remains a significant challenge. This paper introduces ST-Hyper, a novel framework designed to model these dependencies simultaneously. By integrating a spatial-temporal pyramid with adaptive, sparse hypergraph modeling, ST-Hyper achieves state-of-the-art performance, superior noise resilience, and high memory efficiency compared to existing benchmarks.

1. Introduction

Predicting future values for multiple correlated variables—such as traffic speeds across a city or temperature readings across a continent—requires modeling two distinct types of dependencies: spatial (inter-variable correlations) and temporal (intra-variable changes over time).

Traditional models often treat these scales independently or in parallel. For instance, a model might analyze city-level spatial patterns and daily temporal patterns as separate entities. However, real-world data frequently contains patterns that only emerge at specific joint spatial-temporal scales (ST-scales). A local sea-land breeze represents a small-spatial, short-temporal pattern, whereas regional seasonal rainfall represents a medium-spatial, long-temporal pattern. Current methods often fail to capture these "high-order" interactions—relationships involving groups of variables across different scales—leading to sub-optimal predictions, particularly when data is noisy or patterns are heterogeneous.

2. The ST-Hyper Framework

ST-Hyper addresses these limitations through three primary architectural components designed to extract and propagate features across scales.

2.1 Spatial-Temporal Pyramid Modeling (STPM)

The STPM module extracts features using a "fine-to-coarse" approach. It employs Spatial Pyramidal Graph Learning to group correlated variables into clusters, effectively reducing redundancy. Simultaneously, Temporal Multi-Scale Networks utilize 1D Convolutional Neural Networks (CNNs) and pooling layers to capture patterns at varying time resolutions. This pyramidal structure ensures that the model considers both granular and aggregate data views.

2.2 Adaptive Hypergraph Modeling (AHM)

While standard graphs connect two nodes (pairwise), a hypergraph uses "hyperedges" to connect multiple nodes simultaneously. This allows the model to capture group-wise, high-order dependencies. ST-Hyper learns this hypergraph structure adaptively and enforces sparsity. By focusing only on the most robust and meaningful correlations, the model avoids the pitfalls of dense graph models that often overfit to spurious noise.

2.3 Tri-phase Hypergraph Propagation

To facilitate communication between features at different scales, ST-Hyper utilizes a three-step propagation process:

  1. Nodes to Hyperedges: Local features are aggregated into group representations.
  2. Hyperedges to Hyperedges: Graph Attention Networks (GATs) model the interactions between different groups.
  3. Hyperedges to Nodes: Individual variable features are updated with the refined global context.

3. Experimental Evaluation

3.1 Datasets and Baselines

The model was evaluated on six diverse, real-world MTS datasets, each containing over 100 variables to ensure complexity:

  • Traffic: METR-LA and PEMS-BAY (traffic speed).
  • Environment: China-AQI (Air Quality Index).
  • Energy: Electricity consumption and Solar-Energy generation.
  • Weather: A large-scale Temperature dataset featuring 3,850 variables.

ST-Hyper was compared against state-of-the-art baselines, including iTransformer, PatchTST, and CrossGNN.

3.2 Key Results

ST-Hyper consistently outperformed all baselines across the tested datasets:

  • Accuracy: The model achieved an average MAE reduction of 3.8% for long-term forecasting and 6.8% for short-term forecasting.
  • Noise Resilience: In robustness tests using Gaussian white noise (40dB SNR), ST-Hyper maintained its performance lead, outperforming baselines by up to 20.2% in MSE.
  • Efficiency: On the large-scale Temperature dataset, ST-Hyper demonstrated superior scalability. It utilized significantly less GPU memory (~12.8GB) compared to iTransformer (~19.6GB) and Corrformer (~17.2GB) while delivering higher accuracy.

4. Discussion and Insights

The success of ST-Hyper stems from two core insights. First, joint scale modeling allows features from a "city-hour" scale and a "region-day" scale to be treated as nodes in the same hypergraph, enabling them to reinforce each other. Second, sparsity acts as a shield; by selecting only the top-$K$ strongest connections, the model filters out the "spurious" correlations that often plague dense models in noisy environments.

Furthermore, the spatial pyramid coarsening reduces the number of nodes processed at higher levels, allowing the framework to handle thousands of variables (as seen in the Weather dataset) with a manageable memory footprint.

5. Conclusion and Future Work

ST-Hyper represents a significant advancement in multivariate time series forecasting by effectively modeling high-order, multi-scale dependencies. While the current framework is robust, future research will explore dynamic hypergraphs that evolve in real-time during inference and the introduction of directed hyperedges to better model causal relationships (e.g., how an event in one city causes a downstream effect in another). The authors also aim to further optimize the architecture for ultra-large-scale datasets with even higher dimensionality.

Summary updated

BibTeX
@inproceedings{wu2025st,
  author = {Binqing Wu and Jianlong Huang and Zongjiang Shang and Ling Chen},
  title = {{ST-Hyper: Learning High-Order Dependencies Across Multiple Spatial-Temporal Scales for Multivariate Time Series Forecasting}},
  booktitle = {ACM International Conference on Information and Knowledge Management},
  year = {2025},
  doi = {10.1145/3746252.3761281},
}
Figure