Higher-Order Spatio-Temporal Neural Networks for Covid-19 Forecasting

Yuzhou Chen, Sotiris Batsakis, H. Vincent PoorICASSP 2023 · DOI

Abstract

Accurate forecasting of infectious disease spread, such as COVID-19, is critical for public health planning. Traditional deep learning approaches primarily utilize Graph Neural Networks (GNNs) to model spatial dependencies. However, these models are often restricted to pairwise interactions between nodes, failing to capture the complex, multi-way relationships inherent in regional population movements. This paper presents Higher-Order Spatio-Temporal Neural Networks (HOST-NETs), a novel framework that employs simplicial complexes to represent higher-order topological structures. By integrating Hodge-k Laplacians and adaptive graph convolutions with temporal modeling via Gated Recurrent Units (GRUs), HOST-NETs achieve superior predictive performance. Evaluations on the COV19-STG dataset show significant improvements over existing state-of-the-art models.

Introduction

The COVID-19 pandemic highlighted the need for robust spatio-temporal forecasting models. Most existing models treat geographical regions (e.g., counties) as nodes in a graph, with edges representing connections. While GNNs have advanced this field, they are fundamentally limited by their focus on dyadic (pairwise) relationships. In reality, viral transmission often occurs within tightly-knit clusters or "cliques" involving multiple regions simultaneously—such as tri-state metropolitan areas. To address this, we propose moving beyond simple graphs toward simplicial complexes, which allow for the modeling of higher-order interactions.

Methodology

Simplicial Complex Representation

The core innovation of HOST-NETs is the use of simplicial complexes to represent the connectivity of U.S. counties. Unlike a standard graph, a simplicial complex includes higher-dimensional components:

  • 0-simplex: A single node (e.g., a county).
  • 1-simplex: An edge connecting two nodes.
  • 2-simplex: A "filled-in" triangle representing a three-way interaction.
  • 3-simplex: A tetrahedron representing a four-way interaction.
Overview of simplicial complexes from 0-simplex to 3-simplex.

By processing information across these different dimensions, the model can capture community-level dynamics that are invisible to standard node-to-node message passing.

HOST-NET Architecture

The HOST-NET architecture processes spatio-temporal data through several specialized layers. It utilizes Hodge-k Laplacians, a mathematical generalization of the graph Laplacian, to perform convolutions on higher-order structures like edges and triangles. Furthermore, the model employs an Adaptive Spatial Graph Convolution mechanism. Instead of relying solely on a static, pre-defined geographic graph, HOST-NET learns a hidden adjacency matrix from the data, allowing it to adapt to changing social behaviors and policy shifts.

The HOST-NET architecture, illustrating the HOST layer, temporal processing with GRUs, and the final prediction stage.

The spatial features extracted from the simplicial complexes are combined with Gated Recurrent Units (GRUs) to capture temporal trends. An Attention Mechanism is then used to weigh the importance of different connectivity levels (e.g., nodes vs. triangles) for the final prediction.

Datasets

The model was evaluated using the COV19-STG (Spatio-Temporal Graph) datasets, which cover three major U.S. regions from January 2020 to September 2021:

  • East Coast: 668 counties across 14 states.
  • West Coast: 131 counties across 4 states.
  • Central States: 1,944 counties across 21 states.

These datasets integrate time-varying features (daily cases and deaths) with static demographic and behavioral data (e.g., mask-wearing prevalence). Edges were defined based on a driving time threshold of 0.5 hours.

Results and Discussion

HOST-NETs consistently outperformed eight state-of-the-art baselines, including DCRNN and AGCRN. On the West Coast dataset, the model achieved a Root Mean Squared Error (RMSE) improvement of 136.00 points over the next-best baseline.

Ablation studies confirmed the critical role of higher-order structures. Removing the 2-simplex (triangle) convolution led to a significant increase in error rates across all datasets. This suggests that the "triangle" structures represent essential predictive signals—likely corresponding to regional hubs or commuter clusters where transmission dynamics are most intense.

Conclusion

HOST-NETs represent a significant advancement in epidemic forecasting by incorporating higher-order topological information. The ability to model multi-way interactions through simplicial complexes provides a more holistic view of disease spread. Future work will focus on scaling these methods to larger global networks and exploring even higher-order motifs beyond 3-node clusters.

Summary updated

BibTeX
@inproceedings{chen2023higher,
  author = {Yuzhou Chen and Sotiris Batsakis and H. Vincent Poor},
  title = {{Higher-Order Spatio-Temporal Neural Networks for Covid-19 Forecasting}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  doi = {10.1109/icassp49357.2023.10095012},
}
Figure