Heterogeneous Sheaf Neural Networks

Abstract

We introduce HetSheaf, a framework that uses cellular sheaves to handle heterogeneous graphs. Rather than relying on specialised architectural components, our approach encodes heterogeneity through type-aware local feature spaces and learned restriction maps. We also present SheafPool, a readout mechanism for graph-level predictions that maintains invariance to basis changes. HetSheaf achieves performance gains of up to +2 percentage points on the Heterogeneous Graph Benchmark while reducing parameter counts by up to 10× compared to both homogeneous and heterogeneous baselines.

Key Contributions

• A cellular-sheaf framework for heterogeneous graphs (multiple node and edge types) without task-specific architectural components.
• Type-aware restriction maps that encode relational structure directly in the sheaf topology.
• SheafPool — a basis-change-invariant graph-level readout compatible with any sheaf GNN backbone.
• Strong empirical results: up to +2 pp on HGB benchmarks with up to 10× parameter reduction.

Resources

• 📄 ArXiv: https://arxiv.org/abs/2409.08036
• 🧾 BibTeX: https://arxiv.org/bibtex/2409.08036

Recommended citation: Braithwaite, L.; Borgi, A.; Onorato, G.; Tarantelli, K.; Restuccia, F.; Silvestri, F.; Liò, P. (2024). "Heterogeneous Sheaf Neural Networks." arXiv:2409.08036.
Download Paper | Download Bibtex