acceptodds
Under review as a conference paper at ICLR 2027

Hierarchical Pooling for Sheaf Neural Networks

Abstract

Sheaf neural networks (SNNs) extend graph neural networks (GNNs) by operating on cellular sheaves, which assign to each node and edge its own vector space, along with linear restriction maps from node spaces to the spaces associated with their incident edges. These maps determine the sheaf Laplacian and can be learned at each layer, enabling layer-specific diffusion dynamics. Existing SNN architectures, architectures, however, can only operate at a fixed graph resolution. We introduce Hierarchical Sheaf Pooling (HiSP), a pooling layer that enables hierarchical representation learning in SNNs. Following the Select–Reduce–Connect framework, HiSP groups fine-scale nodes, reduces their features using the fine-scale sheaf structure, and constructs the corresponding coarse graph. A new sheaf is then learned at the coarse resolution. By interleaving sheaf diffusion and pooling, HiSP enables hierarchical SNNs to learn distinct diffusion dynamics across layers and graph resolutions. This construction supports multiscale problems in which aggregation changes both the interacting entities and their relations.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
sheaf neural networks, graph neural networks, graph pooling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_QYyFOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Hierarchical Sheaf Pooling (HiSP), a pooling operator for neural sheaf diffusion. Following the Select–Reduce–Connect framework, it partitions the graph into connected clusters, constructs regularised transport maps from the learned restriction maps along rooted spanning trees, and aggregates node features after mapping them into a common reference stalk. A new sheaf Laplacian is learned on the resulting coarse graph. The authors also introduce a sheaf-aware readout for graph-level prediction.

Critical strengths
  • The problem of aggregating sheaf-valued signals across graph resolutions is interesting and relevant to hierarchical graph representation learning. The transport-based reduction is a concrete, mathematically motivated way to make features in different stalks comparable.
  • The construction is explained in reasonable mathematical detail, including regularised least-squares transport, non-expansive normalisation, and an explicit discussion of path and root dependence.
  • Several controlled experiments show substantial gains over unpooled neural sheaf diffusion.
Critical weaknesses
  1. The methodological novelty and necessity of the proposed complexity are insufficiently established. The paper combines standard graph coarsening, established neural sheaf diffusion, regularised inverse maps and a new readout; the specific advantage of sheaf-based transport over simpler aligned or learned feature reductions remains unclear.
  2. The choice of baselines is too limited to substantiate claims about hierarchical learning. On several tasks, the comparison is principally with unpooled NSD and GCN/GIN equipped with EigenPooling. Comparisons against a broader range of hierarchical GNN pooling methods, and especially stronger controls using the same selection and readout but simpler reductions, would be important.
  3. Several synthetic tasks are constructed around the intended hierarchical mechanism. In HierBI, the four connected rooms and the target number of clusters are built into the pooling setup, making it difficult to separate the effectiveness of transport from a favourable prescribed hierarchy. The study does not sufficiently test robustness to other graph structures or imperfect cluster assignments.
  4. The empirical benefit on standard datasets is mixed, which is particularly important for assessing the need for hierarchical pooling. In Table 6, adding HiSP pooling to NSD with the proposed HiSP readout improves MUTAG (79.9±11.2 to 82.5±5.1) and PROTEINS (69.9±4.2 to 74.4±4.1), but not ENZYMES (41.0±5.7 to 37.8±6.4) or NCI1 (74.6±4.1 to 74.2±3.0). Given the variability, the evidence that the pooling step itself yields a reliable improvement is limited. Node-classification results are likewise largely comparable to the unpooled network, rather than consistently better.
  5. The reduction depends on choices of cluster roots, spanning trees, and regularisation/normalisation parameters. While the paper acknowledges root and path dependence (Section 3.3.3 and Section 5), it provides insufficient experimental evidence of predictive stability. Ideally, variability across these choices should be quantified and compared with variation across random seeds and data splits. Moreover, non-expansive edge transports may attenuate contributions from nodes far from the root, introducing a potentially consequential dependence on cluster geometry.
  6. The real-world VoteView task still relies on a constructed topology and predetermined community structure, limiting how strongly it validates generalisation to naturally occurring hierarchies. The construction of the vote-specific cutting point used in node features should also be clarified, including what information is available at prediction time.

There are also minor but worthwhile presentation issues. Section 4.1 contains awkward or insufficiently precise claims (for example, that “the label lives at a coarser sheaf resolution” or that a fine sheaf has “no place to represent the coarse constraint”). Several experimental captions could better define the setup and notation, such as the meaning of boldface in Table 4 and the interpretation of pooling “stages”. The conclusion mixes implementation details, architectural descriptions and experimental claims rather than succinctly identifying the main findings and limitations; its organisation could be improved. Some terminology, such as “distractor-rich regions” and “branch embedding”, would also benefit from clarification.

Overall, the current empirical study does not yet provide sufficient evidence of the robustness, reliability, and generalisability of the proposed method to other datasets where hierarchy is important. The strongest improvements concentrate on controlled synthetic tasks with favourable coarsening structures; broader comparisons and quantitative sensitivity analyses would be needed to justify the method's additional complexity.

Questions for authors
  1. Can the authors isolate the contribution of sheaf-based transport from the benefits of grouping nodes and using the new readout? In particular, does alignment through restriction-map transports outperform mapping each stalk into a shared feature space using learnable projections, followed by pooling, under otherwise matched architectural and training choices? Such a comparison should extend beyond HierBI to multiple datasets.
  2. Why was EigenPooling selected as the principal hierarchical GNN baseline? Could a broader comparison with established graph pooling methods be provided, especially on LRIM, VoteView, and standard graph-classification benchmarks?
  3. Section 3.3.3, lines 239–241: the statement that two different paths can connect a cluster root to the same vertex is potentially confusing, since a fixed spanning tree contains a unique path between them. More importantly, how do accuracy and variance change across root choices, spanning trees, tree resampling, transport path length? Can this sensitivity be compared quantitatively to variance across seeds or data splits?
  4. For HierBI, how much prior structural information is given to SELECT (including the prescribed K=4)? How are the hidden signs assigned, and does performance persist when clusters do not coincide with the four planted rooms? Could the authors test less constrained or noisier hierarchical structures?
5 comments
AuthorsOpenReview ↗

Response QYyF Weaknesses 1,2,3 and 4

We thank Reviewer QYyF for the detailed methodological feedback and presentation suggestions. Below we address each weakness and question point by point, referencing the corresponding sections and tables in the revised manuscript.

Weakness 1 — Sheaf-aware hierarchical pooling

HiSP is the first sheaf-aware hierarchical pooling operator for SNNs. It introduces a new architectural capability: using learned sheaf relations to construct coarse entities, then learning new relations between those entities at the next resolution. Existing sheaf diffusion layers and global readouts do not provide this hierarchy.

The methodological novelty is deriving the reduction from the learned restriction maps, transporting node representations into a common cluster stalk before aggregation. Standard clustering and regularized inversion are building blocks used to define this previously missing operation.

Its empirical contribution is also isolated: with the partition and readout held fixed, replacing ordinary averaging with HiSP raises significantly improve performance in tasks such HierBi, which demonstrates the value of the new reduction beyond the choice of clusters or readout.

Weakness 2 — Choice of baselines

We have added a broader set of hierarchical pooling baselines to the revised version.

The poolers are EigenPool, κ\kappa-MIS, TopKPool, SAGPool, MinCutPool, and DiffPool, on GCN, GIN, and NSD. The readouts are a global mean, SheafPool, and the HiSP readout. The full grids are in Section 4 and in Appendix D.

We also add two stronger sheaf backbones, BuNN and CSNN, each with and without HiSP, on LRIM-16-hard (Section 4.4, Table 4; Appendix D.8, Table 25). HiSP still improves both at the longest tested depth.

The control that keeps the selection and the readout fixed, and simplifies only the reduction, is also in the revision. The partition and the decoder stay fixed. Sheaf transport is replaced by a coordinate-wise mean or by one shared MLP that does not use the restriction maps. That comparison is on HierBI, Marked PathCopy, and VoteView (Appendix D.6, Table 19). Results show that sheaf-aware transport is necessary when the task needs that alignment, which is the setting in which a sheaf network is needed in the first place.

Weakness 3 — Given hierarchy, cluster count, and imperfect clusters

The concern is that HierBI builds the four rooms and the target number of clusters into the setup, so the transport cannot be separated from a favourable prescribed hierarchy, and that other cluster counts and imperfect assignments are not tested.

The planted hierarchy is not given to SEL. It receives only the unlabeled adjacency matrix A\mathbf{A}. Node indices are randomly permuted. Room labels, corridor indicators, and the hidden signs are never input features (Section 4.1 and Appendix D.2). Setting K=4\mathrm{K}=4 chooses a count. It does not assign nodes to the four rooms.

The number of clusters is not fixed at that count. κ\kappa-MIS (Bacciu et al., 2023) sets K\mathrm{K} from the graph and is not told that there are four rooms. At κ=3\kappa=3 it matches the run that was given K=4\mathrm{K}=4 (Appendix D.6, Table 21).

Furthermore, those κ\kappa-MIS clusters are imperfect relative to the planted rooms: they split rooms and corridors instead of recovering the four rooms. Sheaf transport still carries the label on that assignment. A standard pooler on the same imperfect partition does not. The result is the transport under a topological partition, including a partition whose size was not prescribed and a partition that misses the planted rooms.

Weakness 4 — When the pooling step helps

The concern is that the gain on standard datasets is mixed, so the pooling step itself may not be a reliable improvement, and that node classification stays close to unpooled NSD rather than consistently beating it.

HiSP makes interactions between groups of nodes easier to learn. Pooling also compresses information, so its benefit depends on whether those interactions outweigh losing fine detail. This is particularly relevant to node classification, where fine-resolution skip features may already provide sufficient predictive information, leaving limited benefit from the coarse branch (Appendix D.7, Table 24).

The benefit of hierarchy is clear on LRIM. With the same sheaf parameterization and total layer budget, two-stage HiSP improves over NSD at every tested depth (Section 4.4, Table 4).

On standard graph classification, the results show a task-dependent benefit (Section 4.6, Table 6). MUTAG and PROTEINS improve when the pooling stage is added on top of the HiSP readout. On ENZYMES and NCI1 the gain over SheafPool is the readout, and a pooling stage is not required. The mixed standard-benchmark results clarify where this additional hierarchy currently offers value.

AuthorsOpenReview ↗

Response QYyF Weaknesses 5 and 6

Weakness 5 — Root choice and tree sampling

On HierBI we ablate the two choices. Spanning-tree sampling, with the root fixed at the maximum-degree hub, is Appendix D.6, Table 18:

Tree samplingTest accuracy (%)
Random BFS (default, M=1M=1)93.5±1.793.5 \pm 1.7
Deterministic BFS92.7±2.592.7 \pm 2.5
Tree ensemble (M=5M=5)94.2±1.894.2 \pm 1.8

Root selection, with random BFS held fixed, is Appendix D.6, Table 20:

Root selectionTest accuracy (%)
Max-degree root (default)93.5±1.793.5 \pm 1.7
Min-degree root78.2±19.078.2 \pm 19.0
Random root per cluster49.9±2.249.9 \pm 2.2

Section 3.2 fixes the root as the maximum-degree hub in each cluster. The hub is read off the graph, so the rooted trees are built once and cached with the partition (Section 4.8). The three tree samplings stay within about one point. On the trained HierBI checkpoints we keep the hub root and redraw 1010 random BFS trees per test graph. The mean of those per-checkpoint standard deviations of test accuracy is 0.350.35 percentage points, an order of magnitude below the inter-seed variation of 1.71.7–2.52.5. That sentence is in Appendix D.6.

Weakness 6 — VoteView

The four wings are how the graph and the label are built. They are not fed as clusters. SEL is the connected partition of the adjacency with K=4\mathrm{K}=4 (Section 4.5, Appendix D.3); the wings are not read by that selector. Features are xi=di−m\mathbf{x}_i = \mathbf{d}_i - \mathbf{m}, and ballots are labels only.

On the remaining presentation points:

  1. In Section 4.5 and Appendix D.3, we clarified that in the NOMINATE spatial voting model (Poole & Rosenthal, 1985), each senator ii has a static 2D ideological coordinate di∈R2\mathbf{d}_i \in \mathbb{R}^2 for a given Congress, and each roll-call bill has a 2D cutting midpoint m∈R2\mathbf{m} \in \mathbb{R}^2 representing the ideological midpoint of the bill topic under debate. At prediction time, vertex features are xi=di−m∈R2\mathbf{x}_i = \mathbf{d}_i - \mathbf{m} \in \mathbb{R}^2 (with a fraction 1−c1-c of senators randomly negated in Table 5); individual senator ballots are never used as input features, and the model must infer the four-wing coalition pattern solely from noisy ideological positions relative to the bill topic.
  2. In Section 4.1, we replaced the phrasing “the label lives at a coarser sheaf resolution” with a direct statement that the graph label depends on coarse inter-community relationships that differ from the fine intra-community relationships on the input edges, and specified the exact K8,8K_{8,8} room and length-33 corridor construction.
  3. In Section 4.4 (Table 4) and Appendix D.8 (Tables 25–26), we explicitly stated the pooling configuration and the boldface convention.
  4. In Sections 3.6 and 4.3, we defined the three per-cluster descriptor branch embeddings (za(Gram),za(mean),za(stat)\mathbf{z}_a^{(\mathrm{Gram})}, \mathbf{z}_a^{(\mathrm{mean})}, \mathbf{z}_a^{(\mathrm{stat})}) and the on-path wrong-class distractor nodes at first appearance.
  5. The organization of Section 4.5 was improved.
AuthorsOpenReview ↗

Response QYyF Questions

Question 1

To isolate the contribution of sheaf-based transport from the grouping of nodes and from the new readout, the partition and the decoder are held fixed. Sheaf transport is replaced by one shared residual MLP that maps every node into one coordinate space before pooling. The MLP does not use the restriction maps. Its last layer starts at zero, so at initialization the pooling step is a coordinate-wise mean, and training is free to leave that mean. The comparison is on HierBI, Marked PathCopy, and VoteView (Appendix D.6, Table 19).

AlignmentHierBIMarked PathCopyVoteView
Shared MLP48.0±3.348.0 \pm 3.365.5±11.765.5 \pm 11.787.6±1.487.6 \pm 1.4
Damped pseudoinverse (τra→v+\tau_{r_a\to v}^+)92.4±2.092.4 \pm 2.082.7±4.882.7 \pm 4.886.4±2.486.4 \pm 2.4
Adjoint transport (τra→v∗\tau_{r_a\to v}^*)93.5±1.793.5 \pm 1.787.5±4.687.5 \pm 4.688.5±1.988.5 \pm 1.9

On HierBI and Marked PathCopy the clusters are the same in every row, so the gap is the alignment. Adjoint transport uses the learned restriction maps; the shared MLP does not. On VoteView the features are already positions in one Euclidean space, and the shared MLP reaches the same accuracy as adjoint transport. Sheaf-aware transport is necessary when the task needs that alignment, which is the setting in which a sheaf network is needed in the first place.

Question 2

In the revised version we compare multiple poolers (EigenPool, κ\kappaMIS, TopKPool, SAGPool, MinCutPool, and DiffPool) on GCN, GIN, and NSD. The readouts are a global mean, SheafPool, and the HiSP readout. The full grids are in Section 4 and in Appendix D. On LRIM-16-hard the broadened comparison is NSD, BuNN, and CSNN, each with and without HiSP (Section 4.4, Table 4; Appendix D.8, Table 25).

Question 3

In Section 3.3.3 the sentence now reads: “Within a cluster, two different paths π1\pi_1 and π2\pi_2 stemming from distinct spanning trees may connect the root rar_a to the same node vv.” About root choice and tree sampling, see Weakness 5.

Question 4

See Weakness 3.

Reviewer_tz4BOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a hierarchical pooling operation (HiSP) for Sheaf Neural Networks (SNNs). Unlike traditional GNNs, where node representations reside in a shared space, SNNs define representations for each node within distinct stalk spaces. This necessitates a specialized approach to pooling, which explicitly defines how representations from different stalk spaces are related, aggregated, and represented across the graph. The authors propose a sheaf-aware hierarchical pooling operator HiSP. The HiSP layer proceeds by partitioning nodes into clusters (SELECT), aligning and aggregating features within each cluster to derive supernodes representations (REDUCE), and establishing edges between these supernodes to construct a new, coarsened graph (CONNECT). On this coarsened graph, the model learns new restriction maps, allowing each hierarchical level to possess diffusion dynamics tailored to its specific granularity. Finally, a READOUT module compresses the resulting coarsened graph into a global graph representation. Experimental results demonstrate that this method outperforms both traditional graph pooling techniques and SNNs without pooling across various synthetic and semi-synthetic tasks designed by the authors. And It achieves performance comparable to or slightly better than baselines on node and graph classification datasets.

Critical strengths
  1. The paper addresses a core challenge in SNN pooling, the inability of pooling operations to directly aggregate features across different stalk spaces. The design of aligning node features into a shared reference space along intra‑cluster spanning trees before aggregation is well‑motivated. By leveraging regularized matrix inversion and operator‑norm clipping, this method avoids numerical explosion induced by long‑path matrix multiplication, which is innovative.
Critical weaknesses
  1. The method shows limited performance gains on real-world datasets, with its advantages primarily evident on synthetic datasets. Consequently, it is debatable whether incorporating such a complex pooling mechanism is justified for practical applications.
  2. The paper lacks a systematic evaluation of the method’s computational overhead and runtime performance. HiSP involves operations such as matrix inversion, spectral norm computation, and the multiplication of multiple matrices, resulting in significantly higher overall computational complexity compared to conventional pooling operations.
Questions for authors
  1. HiSP involves operations such as spanning tree construction, matrix computation, and path feature alignment. Compared to NSD, GCN, GIN, and standard graph pooling methods, what is the increase in runtime and GPU memory consumption for HiSP?
  2. The paper mentions that the propagation results depend on the paths within the spanning tree and the choice of the root node. How much do different spanning trees and root choices affect the final performance?
  3. How are the number of clusters and pooling ratio chosen in experiments? Are there any parameter robustness results for these hyper‑parameters?
  4. Why is there no significant performance improvement on real-world node classification datasets? Could you provide an analysis of the underlying reasons?
  5. The baseline comparisons have limitations. The selected traditional GNNs and hierarchical pooling layers, specifically GCN and GIN, are quite basic, and the study lacks comparisons with more state‑of‑the‑art methods.
4 comments
AuthorsOpenReview ↗

Response tz4B Weaknesses

We thank Reviewer tz4B for the positive assessment of our problem formulation, tree-alignment design, and non-expansive transport regularization. Below we address each weakness and question point by point, referencing the corresponding sections and tables in the revised manuscript.

Weakness 1 — Where the hierarchy helps

HiSP makes interactions between groups of nodes easier to learn. Pooling also compresses information, so its benefit depends on whether those interactions outweigh loosing fine detail. This is particularly relevant to node classification, where fine-resolution skip features may already provide sufficient predictive information, leaving limited benefit from the coarse branch.

The benefit of hierarchy is clear on LRIM. With the same sheaf parameterization and total layer budget, two-stage HiSP improves over NSD at every tested depth.

On standard graph classification, the results show a task-dependent benefit. HiSP enables effective learning across sheaf resolutions, with substantial gains in the tested long-range and multiscale settings. The mixed standard-benchmark results clarify where this additional hierarchy currently offers value.

Weakness 2 — Runtime and memory

Although HiSP computes regularized edge inversions and operator norms, three structural properties keep its computational overhead small (Section 4.8 and Appendix A):

  1. Stalk dimensions in SNNs are small (d∈{1,2,3,4}\mathrm{d} \in \{1, 2, 3, 4\}), so each d×d\mathrm{d}\times\mathrm{d} linear system is tiny.
  2. Regularized inversions (ρe,v∗ρe,v+εI)−1(\rho_{e,v}^*\rho_{e,v}+\varepsilon\mathbf{I})^{-1} and spectral norms are computed only on the ∣V∣−K|\mathcal{V}|-\mathrm{K} spanning-tree edges, rather than on all ∣E∣|\mathcal{E}| graph edges.
  3. Because SEL and CON depend only on the static graph topology, cluster assignments, coarse graphs, and sheaf Laplacian index structures are precomputed once before training via SheafNetPretransform (Section 4.8 and Appendix F), giving a per-forward runtime complexity of O((∣V∣−K)d3+∣V∣d2h)O((|\mathcal{V}|-\mathrm{K})\mathrm{d}^3 + |\mathcal{V}|\mathrm{d}^2\mathrm{h}) (Appendix A).

We measured forward time, full training-step time, and peak GPU memory on an RTX 3080. The comparison with GCN, GIN, GCN+EigenPool, and uncached HiSP is Appendix E.3, Table 27 (with Figure 7). Cached HiSP against unpooled NSD is the row that answers the question:

DatasetNSD train step (ms)NSD+HiSP train step (ms)Peak train memory
Chameleon (N=2277N=2277)17.17±0.1817.17 \pm 0.1821.15±1.0521.15 \pm 1.05 (1.23×1.23\times)994.6→902.7994.6 \to 902.7 MB (−9.2%-9.2\%)
Texas (N=183N=183)19.16±1.6719.16 \pm 1.6725.63±1.3325.63 \pm 1.33 (1.34×1.34\times)80.3→81.280.3 \to 81.2 MB (+1.1%+1.1\%)
MUTAG (batch 3232)9.67±0.459.67 \pm 0.4518.16±0.5018.16 \pm 0.50 (1.88×1.88\times)71.5→70.671.5 \to 70.6 MB (−1.2%-1.2\%)
HierBI (batch 3232, N=76N=76)9.93±0.249.93 \pm 0.2421.90±1.7021.90 \pm 1.70 (2.21×2.21\times)103.0→83.4103.0 \to 83.4 MB (−19.0%-19.0\%)

GCN, GIN, and EigenPool remain the faster Euclidean layers; their times are in the appendix table. Without the pretransform, the Chameleon forward pass is 441441 ms, about 40×40\times the cached HiSP step. The memory drop on Chameleon and HierBI is the coarse sheaf: after pooling, diffusion runs on G1G_1 with ∣V1∣≪∣V0∣|\mathcal{V}_1|\ll |\mathcal{V}_0|.

AuthorsOpenReview ↗

Response tz4B Questions

Question 1

See Weakness 2

Question 2

On HierBI we separate the two choices. Spanning-tree sampling, with the root fixed at the maximum-degree hub, is Appendix D.6, Table 18:

Tree samplingTest accuracy (%)
Random BFS (default, M=1M=1)93.5±1.793.5 \pm 1.7
Deterministic BFS92.7±2.592.7 \pm 2.5
Tree ensemble (M=5M=5)94.2±1.894.2 \pm 1.8

Root selection, with random BFS held fixed, is Appendix D.6, Table 20:

Root selectionTest accuracy (%)
Max-degree root (default, hub depth ≤3\le 3)93.5±1.793.5 \pm 1.7
Min-degree root (corridor periphery, depth 55)78.2±19.078.2 \pm 19.0
Random root per cluster49.9±2.249.9 \pm 2.2

Section 3.2 clearly states now that we choose the root as the maximum-degree hub in each cluster. The hub is read off the graph, so the rooted trees are built once and cached with the partition (Section 4.8), and on HierBI every transport path has depth at most 33. On the trained HierBI checkpoints we keep the hub root and redraw 1010 random BFS trees per test graph. The mean of those per-checkpoint standard deviations of test accuracy is 0.350.35 percentage points, an order of magnitude below the seed variation. We added this in Appendix D.6.

Question 3 — Cluster count and pooling ratio

HiSP has three sizing modes (Section 3.2 and Appendix A). In our experiments we never selected the communities by hand. Every run uses one of these modes, or more than one of them: (i) an explicit cluster count K\mathrm{K}, (ii) a pooling ratio r=K/∣V∣r = \mathrm{K}/|\mathcal{V}| for recursive spectral selection, or (iii) κ\kappa-MIS (Bacciu et al., 2023), which sets K\mathrm{K} from the graph and does not take K\mathrm{K} or rr as input.

The count does not have to be the planted number of communities. On LRIM-16-hard, two-stage HiSP beats unpooled NSD across the ratios and the κ\kappa orders we swept, once the depth can cover the range of the task (Appendix D.8, Table 26). On HierBI, κ\kappa-MIS is not told that there are four rooms. At κ=3\kappa=3 it matches the run that was given K=4\mathrm{K}=4, and that K=4\mathrm{K}=4 run stays high as the feature noise increases (Appendix D.6, Table 21).

Selection / Noise Setting on HierBICluster Count K\mathrm{K}Test Accuracy (%)
Connected (K=4\mathrm{K}=4, default, σ=0.5\sigma=0.5)4.04.093.5±1.793.5 \pm 1.7
κ\kappa-MIS (κ=3\kappa=3, automatic K\mathrm{K})8.08.093.5±2.993.5 \pm 2.9
κ\kappa-MIS (κ=2\kappa=2, automatic K\mathrm{K})8.08.080.4±15.780.4 \pm 15.7
Connected (K=4\mathrm{K}=4, higher noise σ=0.75\sigma=0.75)4.04.093.0±2.293.0 \pm 2.2
Connected (K=4\mathrm{K}=4, higher noise σ=1.00\sigma=1.00)4.04.090.7±2.790.7 \pm 2.7

Question 4

See Weakness 1.

Question 5

We have expanded the evaluation to include broader pooling comparisons and stronger sheaf backbones.

Six established poolers are now tested with GCN, GIN, and NSD on HierBI and VoteView (Tables 1, 5 and 7). Comparing poolers within NSD directly tests the added value of sheaf-based aggregation. On HierBI, the best conventional alternative remains below 50%, while HiSP reaches 93.5%. The fixed-partition ablation further isolates the contribution of transport (Table 19).

We also evaluate BuNN and CSNN, each with and without HiSP, on LRIM (Table 4). The effects are smaller and vary with depth, but HiSP improves both at the largest tested depth. These comparisons directly address the reviewer’s concern showing that HiSP can also benefit stronger sheaf architectures, while identifying settings where its additional hierarchy offers limited gains.

Thanks for the authors’ response, which addresses my confusion to some extent. I will retain my score.

I thank the authors for the additional experiments and analyses provided during the rebuttal: runtime and memory measurements, the selection and effect analyses of tree sampling, root selection, and the number of clusters and pooling ratio on the synthetic datasets, as well as the comparison against more baselines. It is noticeable that most of these additional analyses were conducted on synthetic datasets.

I apologize for not clearly explaining the reasons for not raising my score earlier. Considering the overall design and task performance, I think the paper's overall quality does not reach the level that would justify raising my score to weak accept.

My main concern is the practicality of the method. The effectiveness of this method varies significantly across different tasks: both the primary performance improvements and the related analytical experiments are concentrated on synthetic datasets (HierBI, RingTransfer, Marked PathCopy, and LRIM). In contrast, on the standard classification benchmarks (Standard Benchmarks), the improvements are relatively minor (or less significant) and quite limited compared to those on the synthetic datasets for both graph-level tasks (which exhibit large variance, Table 6) and node-level tasks (Table 24). I pointed this out in my Summary and Weaknesses before. I remain concerned about the practical utility and task transferability of this paper.

Reviewer_nCwmOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose Hierarchical Sheaf Pooling (HiSP), a pooling layer for sheaf neural networks within the SRC framework. Nodes are partitioned into connected clusters, each with a root and a spanning tree. Node features are transported to the root stalk through Tikhonov-regularized (and norm-clipped) maps built from the restriction maps, then combined through a learned convex combination, and the coarse adjacency is S⊤ASS^\top A S. A new sheaf is then learned on the coarse graph. They also propose a sheaf-aware readout and evaluate the resulting architecture on synthetic and real-world graph- and node-level tasks.

Critical strengths
  • Creating pooling strategies tailored to sheaf networks (i.e., that leverage their induced transport plans) is a natural direction.
  • The idea of transporting node embeddings to their cluster's root before combination is intuitive and easy to follow.
Critical weaknesses

In order of importance:

  • The motivation for why standard pooling cannot be applied to SNNs is thin. The claim that node features "belong to distinct stalk spaces" (l. 142) does not hold in the setting of Sec. 2.2, where all stalks are Rd\mathbb{R}^d. If the hypothesis is that aggregating them in a shared space is a bad idea, it should be tested, e.g., by comparing against sheaf-based networks combined with state-of-the-art pooling operators. The only related comparison (direct mean in Table 7) is on HierBI, which is constructed so that the mean cancels the signal.

  • The technical contribution seems modest. Of the four steps in HiSP (partition, spanning tree and root, alignment, and S⊤ASS^\top A S), only the alignment step (and subsequent aggregation) depends on the sheaf. Selection and connection are standard topology-based choices.

  • The baselines are rather limited. BuNN (Bamberger et al., 2025) and CSNN (Ribeiro et al., 2026) report stronger results than NSD, especially on long-range tasks, and it would be informative to see whether HiSP still offers gains for these more advanced models. The non-sheaf baselines (GCN/GIN with EigenPooling) are also dated, and I would expect modern models such as Graph Transformers. Finally, the benchmarks are narrow. Tasks from the Long Range Graph Benchmark (e.g., peptides-func and peptides-struct) would better stress the need for long-range communication.

  • The statement that ρe,v\rho_{e,v} may not admit an inverse (l. 193) is true for general maps, but not (almost surely) for diagonal or conformal maps, nor for orthogonal maps, which are the most popular choice. For general maps, the cost is cubic in the stalk dimension per tree edge. The norm clipping is also not well motivated: why is preventing norm growth along paths important? When the restriction maps are invertible, the normalization moves the transport away from what the sheaf entails. For orthogonal maps the clipping seems inactive (Sec. 5).

  • The choice of root seems arbitrary, and I wonder whether it breaks permutation invariance/equivariance. The complexity of selection and spanning-tree construction is not discussed.

  • The readout uses three descriptors per supernode, but no experiment measures the importance of each.

  • Minor: Sec. 2.2 states that restriction maps are produced by an MLP, but many works use GNNs (e.g., Bamberger et al., 2025; Ribeiro et al., 2026). Eq. 5 is a definition but is written as an identity.

Questions for authors
  • What is the complexity of the selection strategies and of the spanning-tree construction?
  • Is HiSP permutation invariant/equivariant? How sensitive is it to the choice of root?
  • Can the authors elaborate on the importance of the norm clipping, e.g., with an ablation?
  • What is the contribution of each of the three readout descriptors?
  • Does HiSP still help stronger sheaf models, such as BuNN or CSNN, especially on long-range tasks?
  • Could the authors specify the HierBI graph formally in the main text, including how rooms and corridors attach? Sec. 4.1 says four rooms connected by corridors, and K2,2K_{2,2} appears only as the coarse graph.
5 comments
AuthorsOpenReview ↗

Response nCwm Weaknesses 1 and 2

We thank Reviewer nCwm for the careful reading and constructive feedback. Below we address each weakness and question point by point, referencing the corresponding sections and tables in the revised manuscript.

Weakness 1 — Standard pooling and stalk dimension

Standard pooling can be applied only when every vertex stalk has the same dimension. A coordinate-wise sum or mean is defined only in that case, and even when this happens, equal dimension does not give the stalks a shared basis. In general two features with dim⁡F(u)≠dim⁡F(v)\dim\mathcal{F}(u)\neq\dim\mathcal{F}(v) cannot be combined at all.

Section 2.2 now states that equal stalk dimension is notation, following Bodnar et al. (2022), Bamberger et al. (2025), and Ribeiro et al. (2026); our constructions do not use that assumption. The operator that remains defined when the dimensions differ, and that uses the learned restriction maps when they match, is the transport τu→v=(ρe,v∗ρe,v+εI)−1ρe,v∗ρe,u\tau_{u\to v}=(\rho_{e,v}^{*} \rho_{e,v}+\varepsilon\mathbf{I})^{-1}\rho_{e,v}^{*}\rho_{e,u} (Section 3.3.1, Equation 8).

To test that aggregating features in a shared space is a bad idea, we added a new experiment in this revision. It replaces coordinate-wise aggregation by a learned map into one common feature space, and it is run on HierBI and on two tasks that do not plant a cancelling sign pattern. The partition and the decoder stay fixed, so every row pools the same clusters. Sheaf transport is replaced by one shared nonlinear residual MLP applied to every node before pooling. The MLP does not use the restriction maps. Its last layer starts at zero, so at initialization the pooling step is the coordinate-wise mean, and training is free to leave that mean if a shared feature space is sufficient. The results are in Appendix D.6, Table 19.

AlignmentHierBIMarked PathCopyVoteView
Shared MLP48.0±3.348.0 \pm 3.365.5±11.765.5 \pm 11.787.6±1.487.6 \pm 1.4
Damped pseudoinverse (τra→v+\tau_{r_a\to v}^+)92.4±2.092.4 \pm 2.082.7±4.882.7 \pm 4.886.4±2.486.4 \pm 2.4
Adjoint transport (τra→v∗\tau_{r_a\to v}^*)93.5±1.793.5 \pm 1.787.5±4.687.5 \pm 4.688.5±1.988.5 \pm 1.9

A learned map into one feature space does not replace transport along the restriction maps. On HierBI and Marked PathCopy the clusters are the same in every row, so the gap is the alignment. Adjoint transport uses the learned restriction maps; the shared MLP does not. On VoteView the senator features are already positions in one Euclidean space, and the shared MLP reaches the same accuracy as adjoint transport. A shared space is a sufficient summary when the features already share a basis, and it is not a sufficient summary when they do not. The damped pseudoinverse is the other sheaf-aware pullback of the same transport, the pseudoinverse of τra→v\tau_{r_a\to v} in place of the adjoint, so it still uses the learned restriction maps. It stays with adjoint transport on all three datasets. Both sheaf-aware alignments support the claim. Discarding the sheaf does not.

Weakness 2 — SEL and CON are not sheaf-aware

In the revised manuscript, we formulate HiSP within the full Select–Reduce–Connect–Lift (SRCL) framework (Grattarola et al., 2022; Sections 2.3 and 3), where three components explicitly depend on the sheaf geometry:

  1. RED (Section 3.3): Tree-based transport into the cluster root stalk and learned gated aggregation.
  2. LIFT (Section 3.5): Dual root-to-node broadcast along the same transport trees for node-level sheaf U-Nets.
  3. READOUT (Section 3.6): Stalk-aware channel Gram, coordinate mean, and intra-cluster sheaf Dirichlet energy descriptors.

SEL and CON are not sheaf-aware. Both are computed from the graph topology alone, and that choice is deliberate. Because the learned restriction maps ρ(ℓ)\rho^{(\ell)} change at every training step, making SEL depend on ρ(ℓ)\rho^{(\ell)} would require recomputing graph partitions and spanning trees at every forward pass. On the other hand, decoupling SEL and CON from the evolving restriction maps, allows our SheafNetPretransform (Section 4.8 and Appendix F) to precompute the coarsening hierarchy, rooted tree supports, and coarse sheaf Laplacian index structures once before training, yielding a 4×4\times to 40×40\times speedup in forward and training step times (Appendix E.3, Figure 7 and Table 27).

AuthorsOpenReview ↗

Response nCwm Weakness 3

Weakness 3 — Baselines and the long-range test

Regarding LRGB (peptides-func, peptides-struct) and Graph Transformers, HiSP is a pooling and lifting operator for sheaf networks, not a global self-attention model. On those peptide tasks the interaction distance is unknown, and the score is confounded by positional encodings and by the depth of the prediction head (Tönshoff et al., 2023; Mathys et al., 2026. A local GCN with a deeper head already matches Graph Transformers such as GPS and BuNN there Bianchi, 2026). We therefore test long-range behavior on LRIM-16-hard and Marked PathCopy, where the range the model must cover is known (Section 4.4, Table 4; Section 4.3, Table 3).

We added standard pooling baselines. The poolers are EigenPool, κ\kappa-MIS, TopKPool, SAGPool, MinCutPool, and DiffPool. The readouts are a global mean, SheafPool, and the HiSP readout. The full grids and results are in Section 4 and in Appendix D.

Whether HiSP still helps the stronger sheaves is the LRIM comparison. On LRIM-16-hard (Section 4.4, Table 4; full one- and two-stage grids in Appendix D.8, Table 25), each unpooled sheaf is paired with the same sheaf plus connected HiSP at r=0.5r=0.5. HiSP-equipped models almost always perform better than their corresponding unpooled one.

Model (r=0.5r=0.5, d=2\mathrm{d}=2)L=4\mathrm{L}=4L=8\mathrm{L}=8L=12\mathrm{L}=12L=16\mathrm{L}=16
NSD (unpooled)−0.68±.03-0.68 \pm .03−0.79±.28-0.79 \pm .28−0.61±.14-0.61 \pm .14−0.80±.19-0.80 \pm .19
NSD + HiSP−0.79±.04-0.79 \pm .04−1.30±.09-1.30 \pm .09−1.43±.06-1.43 \pm .06−1.50±.06-1.50 \pm .06
BuNN (unpooled)−1.18±.05-1.18 \pm .05−1.61±.07-1.61 \pm .07−1.75±.06-1.75 \pm .06−1.81±.05-1.81 \pm .05
BuNN + HiSP−1.32±.06-1.32 \pm .06−1.66±.03-1.66 \pm .03−1.73±.05-1.73 \pm .05−1.87±.04-1.87 \pm .04
CSNN (unpooled)−1.56±.02-1.56 \pm .02−1.79±.02-1.79 \pm .02−1.87±.02-1.87 \pm .02−1.85±.04-1.85 \pm .04
CSNN + HiSP−1.62±.02-1.62 \pm .02−1.78±.03-1.78 \pm .03−1.87±.03-1.87 \pm .03−1.91±.02-1.91 \pm .02

As an additional point, Appendix B records the HiSP transport in closed form for these two sheaves.

AuthorsOpenReview ↗

Response nCwm Weakness 4

Weakness 4 — Inversion and norm clipping

We thank the reviewer for pointing out the imprecision at line 193. We have revised Section 3.3.1 and Appendix B to distinguish three cases.

Orthogonal restriction maps. When both endpoint maps are orthogonal, the inverse of each restriction map is its transpose. The Tikhonov-regularized transport is a product of orthogonal matrices scaled by (1+ε)−1(1+\varepsilon)^{-1}. Its operator norm is therefore strictly below one, so clipping is inactive.

Diagonal and conformal restriction maps. A diagonal map is invertible when all its diagonal entries are nonzero, and a conformal map is invertible when its scale is nonzero. In either case, the inverse can become arbitrarily large as an entry or scale approaches zero. Small target scales can be meaningful in CSNNs. In particular, Proposition 4.1 of Ribeiro et al. shows that a zero target map suppresses the node’s diffusion update. For conformal source and target maps with scales ss and tt, the undamped transport has operator norm s/ts/t. Along a tree path, these ratios multiply, so a ratio r>1r>1 at each of mm steps gives amplification by rmr^m. Tikhonov damping replaces s/ts/t with st/(t2+ε)st/(t^2+\varepsilon), which tends to zero as tt tends to zero for fixed ss. Operator-norm clipping caps each step at one, making the composed transport non-expansive.

General or rectangular restriction maps. A restriction map can be rectangular or rank-deficient. We use a Tikhonov-regularized least-squares solve to define transport in these cases as well, and apply the same operator-norm clipping to control amplification along tree paths.

To quantify the effect of operator-norm clipping across sheaf parameterizations, we added Appendix D.6, Table 23 comparing HiSP with and without non-expansive clipping:

TaskSheaf TypeWith Clipping (Default)Without ClippingΔ\Delta AccuracyLocation in Revised Paper
HierBIdiag (d=2\mathrm{d}=2)93.5±1.793.5 \pm 1.784.8±18.784.8 \pm 18.7−8.7-8.7Appendix D.6, Table 23
HierBIgeneral (d=2\mathrm{d}=2)91.0±5.091.0 \pm 5.091.3±5.691.3 \pm 5.6+0.3+0.3Appendix D.6, Table 23
HierBIbundle O(d)O(\mathrm{d}) (d=2\mathrm{d}=2)91.6±2.591.6 \pm 2.591.9±0.991.9 \pm 0.9+0.3+0.3Appendix D.6, Table 23
Marked PathCopygeneral (d=2\mathrm{d}=2)81.8±9.381.8 \pm 9.363.0±13.363.0 \pm 13.3−18.8-18.8Appendix D.6, Table 23
Marked PathCopybundle O(d)O(\mathrm{d}) (d=2\mathrm{d}=2)84.8±2.384.8 \pm 2.384.8±2.384.8 \pm 2.30.00.0Appendix D.6, Table 23

For orthogonal maps the Tikhonov transport has operator norm (1+ε)−1<1(1+\varepsilon)^{-1}<1, so clipping divides by 11 and does not change the map. Marked PathCopy therefore matches exactly. The HierBI bundle means differ by 0.30.3 points, inside the seed standard deviations (2.52.5 and 0.90.9). That gap is run-to-run variation from computing the norm in one branch and not the other. It is not evidence that clipping alters orthogonal transport.

Finally, regarding the O(d3)O(\mathrm{d}^3) inversion cost for general maps: inversions are performed only on the ∣V∣−K|\mathcal{V}|-\mathrm{K} spanning-tree edges (rather than all ∣E∣|\mathcal{E}| graph edges), and stalk dimensions in SNNs are small (d∈1,2,3,4\mathrm{d}\in1,2,3,4), resulting in only a few milliseconds of overhead per forward pass (Appendix E.3, Table 27).

Comments on the paper

1 comment

We wish to express our deep dissatisfaction with the level of engagement with our rebuttal.

We have responded carefully and point by point to the concerns raised, and revised the manuscript accordingly. Yet the brief follow-up replies leave the central question unanswered: which substantive concerns remain unresolved, and why do they still justify rejection after our responses and revisions?

As authors, we took your concerns seriously and put a lot of effort in addressing them. That effort must be reciprocal. A generic acknowledgement, accompanied by an unchanged negative recommendation and no clear explanation of the remaining objections is disrespectul to our work and defeats the purpose of a scientific discussion.

The ICLR reviewer guidelines explicitly ask reviewers to "State your reasoning and what did/didn’t change your recommendation throughout the discussion phase." (https://iclr.cc/Conferences/2027/ReviewerGuidelines). We ask that this standard be reflected in this discussion.

Specifically, we ask the reviewers to identify which concerns have been resolved, which remain, and why the remaining concerns are sufficiently consequential to justify rejection. If our responses are considered insufficient, the reasons should be stated explicitly so that we can engage with them.

We consider this a serious shortcoming in the review process and ask that it be addressed before the final assessment. We also kindly ask the AC to examine the negative recommendations in light of the answers we provided.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.