Sheaf Convolutional Networks (SCNs)

• Sheaf Convolutional Networks are geometric deep learning models that leverage cellular sheaves to incorporate direction sensitivity and cooperative message passing.
• Cooperative Sheaf Neural Networks employ directed sheaf Laplacians and per-node gating mechanisms to control message broadcast and listening, thereby enabling selective path routing.
• The architecture supports efficient multi-hop diffusion with fine-grained edge-aware transformations, making it effective for asymmetrical and heterophilic data structures.

Sheaf Convolutional Networks (SCNs) are a family of geometric deep learning architectures that extend classical graph neural networks by leveraging cellular sheaves—structures that assign vector spaces and restriction maps to vertices, edges, and higher-order cells of a complex. SCNs generalize message-passing on graphs or hypergraphs to encode local edge-aware transformations, model heterophily, and provide precise control over oversmoothing. This entry focuses on Cooperative Sheaf Neural Networks (CSNNs), a recent class of SCNs designed to deliver node-level cooperative message-passing via directed sheaf diffusion and selective communication (Ribeiro et al., 1 Jul 2025).

1. Mathematical Foundations of Sheaf Convolutional Networks

A cellular sheaf $\mathcal{F}$ on a (directed or undirected) graph $G = (V, E)$ specifies:

• For each vertex $i \in V$, a stalk $\mathcal{F}(i) \cong \mathbb{R}^d$ (vector space);
• For each (directed) edge $ij\in E$, a stalk $\mathcal{F}(ij) \cong \mathbb{R}^d$, and restriction maps:
  • Source restriction: $F_{i\to j}: \mathcal{F}(i) \to \mathcal{F}(ij)$;
  • Target restriction: $$F_{i\leftarrow j}: \mathcal{F}(i) \to \mathcal{F}(ji)$$ (for reverse orientation).

The 0-cochain space is $$C^0(G, \mathcal{F}) = \bigoplus_{i\in V} \mathcal{F}(i)$$. A feature field is a vector $\mathbf{X} = (x_1, ..., x_n) \in \mathbb{R}^{nd}$.

Directed Sheaf Laplacians: For each node $G = (V, E)$0, let $G = (V, E)$1 denote the neighbor set. In the directed, flat-bundle setting ($G = (V, E)$2, $G = (V, E)$3 with $G = (V, E)$4 and $G = (V, E)$5), the sheaf Laplacians are:

• Out-degree:

$G = (V, E)$6

• In-degree:

$G = (V, E)$7

Normalized operators are constructed via block-diagonal degree matrices $G = (V, E)$8, leading to

$G = (V, E)$9

The joint sheaf diffusion operator for CSNNs is

$i \in V$0

which encodes successive diffusion through outgoing and then inbound directed sheaf operators.

2. Cooperative Message-Passing: Modes and Gating

A principal innovation of CSNNs is node-local control over outgoing (broadcast) and incoming (listening) message flow, realized by per-node scaling of the restriction maps:

• Broadcasting is controlled by the scalar in $i \in V$1; setting it to zero blocks all outgoing messages from $i \in V$2.
• Listening is controlled by the scalar in $i \in V$3; setting it to zero blocks all incoming messages to $i \in V$4.

For each node $i \in V$5, possible behaviors are:

• Standard: both scales nonzero; node participates normally;
• Listen-only: outgoing scale zero, inbound scale nonzero; node absorbs but does not transmit;
• Broadcast-only: incoming scale zero, outbound scale nonzero; node transmits but does not absorb;
• Isolate: both scales zero; node becomes communication-invisible (“quarantine”).

At the operator level, if the listening flag for $i \in V$6 is zero, all diffusion terms associated with $i \in V$7 vanish, and the node completely ignores neighbor updates; similarly, if a neighbor’s broadcasting flag is zero, its features are not aggregated.

3. Per-Layer Update Rule and Learning Dynamics

In CSNNs, node features at layer $i \in V$8 are updated by:

$i \in V$9

where:

• $\mathcal{F}(i) \cong \mathbb{R}^d$0 is a learnable layerwise skip connection;
• $\mathcal{F}(i) \cong \mathbb{R}^d$1 is a pointwise nonlinearity (e.g., GELU);
• $\mathcal{F}(i) \cong \mathbb{R}^d$2 and $\mathcal{F}(i) \cong \mathbb{R}^d$3 are weight matrices;
• The per-node broadcast/listen gates and Householder parametrizations are learned by compact neural networks at each layer.

The high-level implementation involves:

• Learning restriction maps using small neural networks (“MapNet”s);
• Building and normalizing in/out Laplacians;
• Composing the joint operator and executing the feature update.

Computational complexity per layer is $\mathcal{F}(i) \cong \mathbb{R}^d$4 for building sparse block-matrices, plus $\mathcal{F}(i) \cong \mathbb{R}^d$5 for applying them, with memory dominated by two sparse Laplacians and per-node maps. Orthogonal parametrizations (Householder reflections) cost $\mathcal{F}(i) \cong \mathbb{R}^d$6 per node.

4. Expressivity: Selective Attention & Oversquashing Avoidance

CSNNs enable two forms of expressive receptive field control:

• Receptive Field Doubling: Each layer’s update can depend on all nodes up to $\mathcal{F}(i) \cong \mathbb{R}^d$7 hops away after $\mathcal{F}(i) \cong \mathbb{R}^d$8 layers, since $\mathcal{F}(i) \cong \mathbb{R}^d$9 propagates across two neighborhoods per layer.
• Selective Path Routing: With appropriate gating, CSNNs can deterministically route a signal selectively from a source $ij\in E$0 to a target $ij\in E$1 across a chosen $ij\in E$2-length path, while ignoring all intervening nodes. This enables reproduction of sequences or chains without interference from side branches—a behavior that mitigates oversquashing noise inherent to traditional GNNs, even in deep networks.

5. Comparison with Classical and Undirected Sheaf Networks

Standard undirected Sheaf Neural Networks (SNNs) employ a symmetric Laplacian and cannot distinguish message direction. Turning off a restriction map disables transmission in both directions. By contrast, CSNNs introduce:

• Edge orientation: Directed edges and decoupled source/target maps;
• Two Laplacians: Separate out- and in-degree Laplacians, breaking symmetry;
• Composed diffusion: The sequential operator $ij\in E$3 enables direction-sensitive, cooperative updates;
• Fine-grained per-node control: Four distinct node communication modes.

These enhancements yield richer modeling of heterophilic and directed structures, faster $ij\in E$4-hop per layer propagation, and intrinsic mitigation of oversquashing—properties unattainable with symmetric, undirected SCNs.

6. Algorithmic Workflow and Implementation

A CSNN layer enacts the following pipeline:

1. Input features $ij\in E$5, graph $ij\in E$6;
2. Learn per-node source/broadcast maps via $ij\in E$7, and target/listen maps via $ij\in E$8;
3. Assemble $ij\in E$9, $\mathcal{F}(ij) \cong \mathbb{R}^d$0;
4. Normalize to $\mathcal{F}(ij) \cong \mathbb{R}^d$1, $\mathcal{F}(ij) \cong \mathbb{R}^d$2;
5. Construct $\mathcal{F}(ij) \cong \mathbb{R}^d$3;
6. Update features through the combined residual and nonlinear pathway.

Batch sparse-dense multiplies and stable Householder-based orthogonal parameterizations (e.g., using PyTorch Geometric and torch-householder) are central to efficient training and inference.

7. Empirical Performance and Scope of Applicability

CSNNs exhibit superior performance to both classical SCNs and prior cooperative GNNs on benchmarks where information flow directionality, selective attention, or heterophily are critical (Ribeiro et al., 1 Jul 2025). This robustness is achieved despite the increased flexibility afforded by dynamic node-level gating and two-step directional diffusion. The architecture is especially suited for domains with strong asymmetry, nonconservative flows, and long-range dependencies, where traditional message-passing saturates quickly, suffers from oversquashing, or fails to capture multi-hop signals.

---

References:

• "Cooperative Sheaf Neural Networks" (Ribeiro et al., 1 Jul 2025).