Sheaf Diffusion with Adaptive Local Structure for Spatio-Temporal Forecasting
Sheaf Diffusion with Adaptive Local Structure for Spatio-Temporal Forecasting
Authors: Abeer Mostafa, Raneen Younis, Zahra Ahmadi Date: 2026-04-13 Paper ID: openalex:2604.11275
Summary
This paper proposes ST-Sheaf GNN, a novel framework for spatio-temporal forecasting that replaces global message passing with information flow over locally structured sheaf-theoretic vector spaces. By utilizing dynamic restriction maps that adapt to local spatio-temporal patterns, the model effectively captures heterogeneous interactions while avoiding the oversmoothing typical of deep graph architectures. Experimental results across multiple real-world benchmarks demonstrate that this sheaf-theoretic approach consistently outperforms conventional GNN models.
Key Contributions
- Introduces ST-Sheaf GNN, a framework that models information flow via dynamic restriction maps within sheaf-theoretic vector spaces.
- Enables expressive modeling of non-intuitive, localized spatio-temporal interactions that are often missed by traditional message-passing GNNs.
- Achieves state-of-the-art forecasting performance by mitigating oversmoothing through the explicit learning of local spatio-temporal topological structures.
Open Questions & Future Work
Key Concepts
- st-sheaf-gnn: A graph neural network framework for spatio-temporal forecasting that models information flow via dynamic restriction maps within sheaf-theoretic vector spaces.
Archivist Review
Approved the central model framework as a reusable concept for sheaf-based topological graph learning and a scoped open question regarding the extension of this theory to heterogeneous graph structures. All other candidates were rejected as either too paper-specific or covered by the selected items. I maintained a high bar for inclusion, ensuring only foundational theoretical contributions that transcend the specific forecasting task were admitted.
Approved Concepts
- Spatio-temporal sheaf diffusion (ST-Sheaf GNN): It provides a novel topological approach to spatio-temporal modeling that addresses the limitations of standard message passing and oversmoothing.
Approved Open Questions
- Sheaf Diffusion for Heterogeneous Graphs: Moving beyond simple graph structures is essential for applying sheaf-based topological modeling to more realistic, high-dimensional, and multi-typed real-world data systems.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.11275
- paper_id
- 2604.11275
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
-
- concept_slugs
- st-sheaf-gnn
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-16T05:46:30Z