Track: Track1; Team name: howyadoin; Model: Neural Sheaf Propagation

[lapry34]

Track

Track 1 — Graph Neural Networks (GNNs)

Team Name

howyadoin

• Dario Loi — Northeastern University
• Gabriele Onorato — Northeastern University

Model

Neural Sheaf Propagation (NSP)

Status

Ready for review

Summary

Track

Track 1 — Graph Neural Networks (GNNs)

Team Name

howyadoin

• Dario Loi — Northeastern University
• Gabriele Onorato — Northeastern University

Model

Neural Sheaf Propagation (NSP)

Status

Ready for review

Summary

This PR contributes a TopoBench-native implementation of Neural Sheaf Propagation (NSP), a generalization of GNNs to cellular sheaves inspired by "Neural Sheaf Diffusion" (Bodnar et al., 2022) and "Surfing on the Neural Sheaf" (Suk et al., 2024). NSP learns representations by propagating signals over a learned sheaf structure where each edge carries a d x d restriction map (diagonal, orthogonal, or general). The model employs a second-order discrete propagation scheme that generalizes standard GNN updates to the sheaf-theoretic setting.

The implementation reuses the nsd_utils ecosystem for sheaf learning, orthogonal parameterization, and Laplacian construction. It introduces three new inductive propagation models (diag, bundle, and general) and a dedicated NSPEncoder wrapper that plugs into the TopoBench pipeline. The bundle variant with orthogonal restriction maps and normalized Laplacians is the default configuration, ensuring stable and expressive transport of signals across the graph.

Implementation

• Inductive discrete propagation models mirroring NSD's stalks-and-maps architecture but utilizing a second-order discrete update rule (topobench/nn/backbones/graph/nsd_utils/inductive_discrete_propagation_models.py).
• Three sheaf structure variants — diag / bundle / general — providing flexibility in restriction map parameterization (diagonal, orthogonal bundle, or full matrix).
• NSPEncoder wrapper integrated into the TBModel pipeline, providing a consistent API for graph and node-level tasks (topobench/nn/backbones/graph/nsp.py).
• Hydra config (configs/model/graph/nsp.yaml) with paper-faithful defaults and support for hyperparameter sweeps (sheaf_type=bundle, d=3, num_layers=2).
• Unit tests covering all variants, stalk dimensions, and sheaf types; ensuring high coverage of the new propagation logic.
• graph/nsp added to the end-to-end pipeline tests in test/pipeline/test_pipeline.py.
• Challenge notebook (2026_tdl_challenge/run_evaluation.ipynb) verified with NSP across the GraphUniverse benchmark, with results included in the challenge submission.

Reference

Bodnar, Di Giovanni, Chamberlain, Liò, Bronstein, "Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs," NeurIPS 2022.

Suk, Giusti, Hemo, Lopez, Barmpas, Bodnar, "Surfing on the Neural Sheaf" NeurIPS Workshop on Symmetry and Geometry in Neural Representations 2022.

• ArXiv (NSD): https://arxiv.org/abs/2202.04579
• OpenReview (Surfing): https://openreview.net/forum?id=xOXFkyRzTlu

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB