acceptodds
Under review as a conference paper at ICLR 2027

TopoU-Net: A U-Net Architecture for Topological Domains

11 views

Abstract

Many modern datasets mix points, edges, regions, groups, objects, events, hyperedges, and relations. Yet neural architectures often force such data into grids, graphs, or sequences, obscuring higher-order structure and making encoder-decoder designs domain-specific. We view U-Net not as a grid-specific architecture, but as a hierarchical encoder-decoder principle: representation spaces, transport maps between levels, and skip connections between matched levels. Combinatorial complexes naturally supply these ingredients through cells, incidences, and ranks. We introduce TopoU-Net, a rank-path U-Net for topological domains. Given a path from an input rank to a bottleneck rank and back, the encoder lifts cochains upward along incidence maps, the decoder transports them downward, and skip connections merge features at matched ranks. Rank replaces spatial scale: choosing paths through nodes, edges, faces, hyperedges, or global cells becomes the central architectural decision. A key quantity is the bottleneck support ratio, the number of cells at the bottleneck relative to the number of cells at the input rank. This ratio is fixed by the complex and chosen path rather than by arbitrary pooling, and it clarifies when skip connections are optional, useful, or structurally important. Across node classification, graph classification, hypergraph node classification, point cloud classification, and image reconstruction, TopoU-Net provides a reusable encoder-decoder template for higher-order structured data. Among the evaluated baselines, it achieves the strongest mean accuracy on three of six node-classification datasets against topological baselines and four of five hypergraph datasets, with the largest gains on heterophilic graphs. Ablations show that removing skip connections is most damaging under severe bottleneck compression.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
topological deep learning, combinatorial complexes, U-Net architectures, higher-order networks, graph neural networks, encoder-decoder architectures, rank-based learning, incidence-based message passing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_deL5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

U-Net is an architecture for image processing that uses an encoder–decoder structure to compress information and then reconstruct detailed outputs, with skip connections preserving fine-scale features. In this paper, U-Net is generalized to combinatorial complexes via rank-wise cell representations, incidence-based transport between ranks, refinement maps, bottlenecks, and skip connections. The paper demonstrates the framework on graph and node-classification, hypergraph, point-cloud, and image-segmentation datasets.

Critical strengths
  • The paper is generally well presented and uses concepts familiar from topological deep learning, making the architecture and exposition easy to follow.
  • The versatility of the architecture is demonstrated across several types of data, including graphs, point clouds, and hypergraphs. For point clouds, the rank structure provides an interesting way to represent different aggregation levels over neighboring points.
  • The hypergraph experiments are particularly relevant, since these datasets are designed to capture higher-order relations. The results suggest that an encoder–decoder architecture with bottlenecks and skip connections may provide a useful alternative to existing hypergraph architectures.
Critical weaknesses
  • The theoretical contribution is relatively limited. Some results might be more appropriately presented as observations or remarks. For example, Proposition 3.5 appears to follow from the composition of compatible equivariant functions.
  • The architecture involves several modelling choices that require further justification. For example, the graph experiments add a single rank-3 cell incident to the triangles. In the image-segmentation experiments, latent higher-level cells are used, but their role and construction could be explained more clearly.
  • The experimental results are mixed across subsections. In some cases, TopoU-Net performs best, but the improvements appear relatively small and are accompanied by substantial variance.
  • The term “canonical instantiation” is somewhat misleading because the paper gives examples of possible refinement and merge functions rather than specifying one concrete implementation.
  • Additional implementation details are needed to make the experiments reproducible.
Questions for authors
  • How far does the analogy of the bottleneck from the classical U-net take us here? For a fixed number of vertices, the number of cells can grow with the rank and there may not be a bottleneck at all. In that case the role of the architecture (and in particular skip connections) is unclear.
  • What is the expressivity? For graphs for example does Topo Unet learn from neighborhoods of cells at any level? (or do they do so only if one uses within-rank refinements based on on incidence-induced adjacencies?) Is the expressivity thus not reduced compared e.g. to GNNs that aggregate over large neighborhoods?
  • More generally, which existing architectures can be represented within the proposed framework? A more formal discussion of this question would help position the contribution relative to existing topological message-passing methods.
5 comments

We thank the reviewer for the careful evaluation and constructive feedback. We appreciate the positive assessment of the presentation, architectural versatility, and hypergraph experiments. We address each concern below.

AuthorsOpenReview ↗

Weakness 1. Theoretical Contribution

Weakness 1. Theoretical Contribution

We agree that Proposition 3.5 follows from the composition of compatible equivariant operators. Its purpose is to establish that the complete encoder--decoder architecture preserves equivariance, rather than to claim a new equivariance principle. Our main contribution is architectural: a rank-indexed encoder--decoder with incidence-based transport, matched-rank skip connections, and an explicit characterization of structural compression. In particular, Proposition 3.6 connects bottleneck dimensions to exact reconstruction without skips. We will clarify this distinction and present Proposition 3.5 as a structural guarantee.

Weakness 2. Modeling Choices

We thank the reviewer for highlighting these modeling choices. The rank-3 cell in the original graph experiments is an explicit global aggregation cell incident to rank-2 triangles, rather than an intrinsic topological feature of every graph. It provides a global-context bottleneck for cross-rank information transport.

To further address the dependence on manually specified rank paths, we introduce a learned soft rank-routing extension that adaptively weights admissible cross-rank transitions using attention while preserving incidence-based message passing and matched-rank skip connections.

Across ten splits, learned routing improves Roman-empire accuracy from 86.39%86.39\% to 88.32%88.32\% and Macro-F1 from 80.91%80.91\% to 83.65%83.65\% relative to validation-selected fixed routing. On Amazon-ratings, accuracy improves from 42.18%42.18\% to 42.97%42.97\% and Macro-F1 from 29.78%29.78\% to 31.06%31.06\%. On Minesweeper, learned routing matches the best fixed-path accuracy (83.98%83.98\%) while improving Macro-F1 from 71.76%71.76\% to 72.91%72.91\%. All fixed-routing controls use the same routing backbone.

Weakness 3. Mixed Experimental Results

We agree that improvements vary across datasets and do not claim uniform superiority. However, the results provide evidence of the framework's effectiveness across several domains. In particular, Table 4 reports the highest mean performance on four of five hypergraph datasets. Moreover, the controlled ablations in Appendix B.3.1 (Table 14) isolate rank-path effects under matched depth and parameter budgets. On Roman-empire at 25% support, the edge-only path achieves 68.56%, whereas the rank-2 path achieves 82.69%. These results indicate that the benefits cannot be explained solely by model size.

Weakness 4. Canonical Instantiation

We appreciate this observation. Section 3.5 specifies incidence-based rank transport and a default additive skip merge, while allowing alternative refinement and fusion operators. This flexibility is intentional: TopoU-Net defines an architectural framework rather than a single fixed network. Nevertheless, we agree that canonical instantiation may suggest a uniquely specified implementation. We will replace it with reference instantiation and explicitly distinguish the general framework from the configurations used in each experiment.

Weakness 5. Reproducibility

We agree that the implementation description can be improved. Appendix B.6 provides common training and architectural settings, but additional dataset-specific information would facilitate reproduction. We will expand the experimental specifications to identify the rank paths, cell-construction rules, incidence normalization, refinement and merge operators, model dimensions, and training configurations used for each dataset.

Question 1. Bottleneck Interpretation

We agree that increasing rank does not necessarily imply monotonic compression. However, combinatorial complexes permit an explicit bottleneck construction: for n0>1n_0>1, introducing a single global cell at a new maximal rank yields ρbot=1/n0<1\rho_{\mathrm{bot}}=1/n_0<1, regardless of intermediate cell counts. Thus, bottleneck formation is a controllable property of the complex construction and selected rank path, rather than an assumption of monotonic downsampling. Section 3.4 formalizes this through support ratios. Table 9 further shows that removing skip connections causes substantially larger performance degradation under severe bottleneck compression. We will clarify this distinction in the revision.

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

TopoU-Net introduces a new U-Net inspired encoder-decoder architecture for combinatorial complexes. Their strategy during the encoder portion consists of pooling from lower dimensional cells to higher-dimensional cells (i.e. pooling from vertex to edge and then from edge to triangle and so on). For the decoder portion, they downsample in much the same way, go from edge to vertex etc. From my understanding, this boils down to propagating messages from higher-order cells (e.g. edges) down to lower-order cells (e.g. vertices). In their experiments, they show the accuracy of TopoU-Net on various graph and node classification benchmarks. They provide two ablations: (1) necessity of various ranks in their rank-path compression and (2) skip connections.

Critical strengths

It is useful to think of U-Nets over combinatorial complexes instead of simplifical complexes. I think it adds a layer of flexibility as simplicial complexes are a special case of the combinatorial complex and you are not constrained strictly to the face-inclusion rules for constructing simplicial complexes. There have already been several papers suggesting that it is sometimes useful to consider higher-order interactions (e.g. https://arxiv.org/abs/2010.03633, https://arxiv.org/abs/2106.12575, ) rather than only edges/vertices. The aggregation path also makes sense for situations in which we are dealing with higher order structures.

Critical weaknesses

My main issue with this manuscript is with the experiments and correspondingly, the novelty of this work. Overall, I am concerned that the experiments do not convincingly establish the benefit of using this kind of topological representation along with a U-Net encoder/decoder.

  1. Part of their experiment is showing the performance of their network on various standard graph benchmarks. However, I am unsure why using a higher-order topological representation for these tasks is helpful. Indeed, it seems like for datasets like ModelNet10, ModelNet40 (Table 5), MUTAG, PROTEINS (Table 3), their performance does not seem to be much better than standard baselines. Furthermore, in the case of graph level classification tasks such as with the MUTAG datasets, I'm even more unsure of how using some kind of higher-order topological representation is helpful as they are doing. There are no graphs in MUTAG that contain 3-cliques so I'm not sure how helpful the 0→1→20 \to 1 \to 2 rank path aggregation is.
    I acknowledge that in the discussion of the results in Table 1, they provide some intuition that perhaps on heterophilic graphs, their Topo-UNet seems to perform better because these datasets require more global information. However, I'm wondering how necessary the aggregation through 3-cycles is? Is it just the global n3n_3 cell that is improving the performance of TopoU-Net?

  2. Speaking of the n3n_3 cell, in the ablations, it looks like it is necessary to improve the performance of TopoU-Net. Therefore, I think it is necessary to compare to baselines which incorporate global structure so we can understand whether the rank-path encoding/decoding is actually improving performance of TopoU-UNet. For instance, an easy way to incorporate global structure is simply to attach a virtual node to each GraphSAGE, GAT, GIN, or H2GCN. There are also graph transformer architectures which allow global information to be propagated to each node. I think this comparison is necessary to understand the utility of TopoU-Net.

  3. Also, it seems that in the experiments the way the combinatorial complexes are all constructed so that they are all simplicial complexes. Is there a scenario where the generalization of a combinatorial complex is necessary?

Finally, I believe the writing and clarity could be improved. While the rank-path formalization is useful for their generalization of U-Nets to combinatorial complexes, I think this paper could be much more understandable to the general ICLR audience if they provide an example of what this means for the case of vertices and edges on the graphs. I believe that this boils down to pooling and un-pooling along edges and generalizing to higher-order cells.

Questions for authors

Could the authors address the my questions from the "Weaknesses" section?

3 comments

We thank the reviewer for the careful evaluation and constructive feedback. We appreciate the recognition of the flexibility of combinatorial complexes and the relevance of higher-order interactions. We understand the concern that the experiments should more clearly establish the benefits and novelty of the proposed rank-based U-Net architecture. We address the three critical weaknesses and the presentation suggestion below.

Weakness 1. Benefits of Higher-Order Representations

We agree that higher-order representations do not necessarily improve every task, and we do not claim uniform superiority over standard baselines. However, the MUTAG concern assumes that rank-2 cells must correspond to triangles.

But, For MUTAG, rank-2 cells are constructed from a cycle basis of each molecular graph, allowing rings of different lengths to be represented as higher-order cells than only 3-cliques. Consequently, cycles containing more than three vertices can be represented as rank-2 cells. The absence of triangles therefore does not imply the absence of rank-2 structure in this construction. We will clarify the distinction between cycle-based and triangle-based lifting in the manuscript.

Regarding the contribution of higher-order ranks, Appendix B.3.1 (Table 14) provides controlled comparisons under comparable depth and parameter budgets. On Roman-empire, at 25%25\% support, the edge-only path achieves 68.56%68.56\% accuracy, whereas the rank-2 path achieves 82.69%82.69\%, despite comparable parameter counts (137,707 versus 137,004). At 50%50\% support, the corresponding accuracies are 71.92%71.92\% and 83.26%83.26\%.

Importantly, these comparisons do not require the global rank-3 cell. They therefore provide evidence that rank-2 incidence-based aggregation can be beneficial independently of global rank-3 communication, and that the observed improvements cannot be explained solely by increased parameter count.

Weakness 2. Global Rank-3 Cell and Global Baselines

We agree that the contribution of rank-structured information propagation should be distinguished from generic global aggregation. To address this concern, we provide three complementary pieces of evidence: capacity-controlled experiments from the original manuscript, new rank-path ablations with global communication retained, and additional virtual-node baselines.

Higher-order aggregation without global rank-3 communication Appendix B.3.1 (Table 14) already isolates the contribution of rank-2 aggregation without introducing the global rank-3 cell. On Roman-empire, at 25% support, the rank-2 path achieves 82.69% accuracy compared with 68.56% for the edge-only path, an improvement of 14.13%percentage points. At 50% support, the corresponding accuracies are 83.26% and 71.92%, an improvement of 11.34 points. Importantly, the models have comparable parameter counts (137,004 versus 137,707 at 25% support). These results show that substantial benefits can arise from rank-2 aggregation even in the absence of global rank-3 communication.

We additionally compare two TopoU-Net configurations that both retain a global rank-3 cell. Full uses vertices, edges, triangles, and connected-component cells, whereas component-only has the same architecture except there are no rank-2 cells or triangles.

:

\text{FULL:}\quad& 0\rightarrow1\rightarrow2\rightarrow3 \rightarrow2\rightarrow1\rightarrow0,\\ \text{Component-only:}\quad& 0\rightarrow1\rightarrow3 \rightarrow1\rightarrow0. \end{aligned}$$ The component-only path provides global aggregation without traversing rank-2 cells, whereas FULL additionally incorporates rank-2 incidence-based transport. **Table: Additional global-information comparisons.** Test accuracy (%, mean ± standard deviation) over 10 seeds. Both TopoU-Net variants include global rank-3 aggregation. | Model | Squirrel | Chameleon | |:---|:---:|:---:| | GCN + virtual node | 27.64 ± 0.91 | 40.88 ± 3.17 | | GraphSAGE + virtual node | 31.37 ± 1.32 | **48.37 ± 1.96** | | GAT + virtual node | 25.45 ± 1.02 | 39.01 ± 2.72 | | GIN + virtual node | 31.50 ± 1.36 | 40.64 ± 1.69 | | H2GCN + virtual node† | 24.80 ± 1.47 | 37.34 ± 2.54 | | **TopoU-Net (component-only)** | 32.83 ± 2.54 | 42.42 ± 2.97 | | **TopoU-Net (FULL)** | **34.06 ± 2.66** | 46.55 ± 2.88 | On Chameleon, FULL achieves 46.55% accuracy compared with 42.42% for component-only, a gain of 4.13 percentage points. On Squirrel, FULL improves accuracy from 32.83% to 34.06%. Since both variants already contain global rank-3 aggregation, these results indicate that the full rank path can provide benefits beyond the global cell alone. Thus, global communication alone does not consistently reproduce the behavior of the full rank path. Taken together, Table 14 demonstrates that rank-2 aggregation can improve performance without a global cell, while the new ablation shows additional gains from the full rank path when global aggregation is already present. These results support the contribution of incidence-structured cross-rank propagation beyond simply introducing a global cell.

Weakness3. Combinatorial Versus Simplicial Complexes

We thank the reviewer for this question. We agree that triangle-based constructions can be represented simplicially. However, TopoU-Net is not restricted to such constructions, and we now provide two concrete non-simplicial examples.

First, in our MUTAG graph-classification implementation, rank-2 cells are cycles constructed using nx.cycle_basis(G), rather than being restricted to triangles. The resulting basis can contain cycles of length greater than three. Such a cycle can therefore be represented as a single rank-2 cell even though it is not a rank-2 simplex. For example, a cycle on k>3k>3 vertices may be assigned

\qquad \operatorname{rk}(c)=2.$$ Thus, the absence of 3-cliques does not imply the absence of rank-2 structure under our cycle-based construction. Second, to make this distinction explicit, we additionally evaluate TopoU-Net on dataset CSL using 4-cycles as rank-2 cells. A four-cycle $$c=\{v_1,v_2,v_3,v_4\}, \qquad \operatorname{rk}(c)=2,$$ is non-simplicial in this rank assignment: a simplex containing four vertices has dimension three, not two. Using the rank path $0\rightarrow1\rightarrow2\rightarrow1\rightarrow0$ with 4-cycles as rank-2 cells, TopoU-Net achieves $18.67\pm9.32\%$ test accuracy over 10 folds. Under the same evaluation setup, the corresponding baselines achieve $8.00\pm2.81\%$ for GCN, $10.00\pm3.51\%$ for GraphSAGE, $7.33\pm2.11\%$ for GAT, $8.67\pm3.22\%$ for GIN, $11.33\pm6.32\%$ for GIN-pool, and $8.67\pm3.22\%$ for Graph U-Net. These experiments demonstrate that the combinatorial-complex formulation is not merely formal generality: the same rank-indexed encoder--decoder can operate on higher-order cells that are not simplices of the corresponding rank. The CSL result provides evidence that a non-simplicial construction can be useful, but we do not claim that such constructions are necessary for every dataset or that this experiment establishes a universal predictive advantage. We will clarify this distinction and the exact cell constructions used in each experiment in the revision.
Reviewer_btXtOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TopoU-Net, a U-Net-inspired architecture for topological deep learning that aims to model data with higher-order relational structures. The authors argue that existing architectures, particularly conventional Graph Neural Networks (GNNs), may be limited in their ability to capture complex higher-order interactions. The central idea is to reinterpret U-Net as a general-purpose encoder–decoder architecture rather than one specifically designed for image grids.

The proposed architecture incorporates skip connections to preserve information across encoding and decoding stages and introduces a bottleneck support ratio to characterize structural compression. The authors evaluate TopoU-Net across graph, hypergraph, and image-related tasks. The experimental results suggest that the approach is promising, with particularly strong performance on hypergraph node classification and heterophilic graph benchmarks.

Critical strengths
  1. S1. Extensive experimental evaluation. The paper evaluates the proposed architecture across a diverse set of datasets and learning tasks, including node classification, graph classification, and hypergraph-related tasks. This broad experimental coverage is valuable for assessing the generality of the proposed framework.
  2. S2. Well-defined problem formulation. The paper presents a clear problem formulation and provides a reasonably structured mathematical description of the proposed architecture.
  3. S3. Promising performance on higher-order relational data. The proposed approach appears particularly effective when modeling higher-order structures, with encouraging results on hypergraph datasets. These findings suggest that the architecture could be a useful contribution to topological deep learning.
Critical weaknesses
  1. W1. Clarity and presentation. The paper would benefit from substantial improvements in writing and overall narrative coherence. In particular, the abstract is dense and does not clearly communicate the motivation, methodological contributions, and main empirical findings as a cohesive story. Several statements also feel overly general and would benefit from more precise technical explanations. I encourage the authors to revise the abstract and introduction to establish a clearer connection between the limitations of existing methods, the proposed architectural choices, and the resulting contributions.
  2. W2. Limited comparison with modern vision architectures. Although the paper positions TopoU-Net as a generalization of the U-Net paradigm and includes image-related experiments, it would be interesting to investigate how the proposed approach compares with more recent vision architectures, particularly Vision Transformers (ViTs) and other transformer-based models. Such comparisons could help clarify the practical advantages of the proposed framework beyond traditional convolutional architectures. If relevant to the paper's broader claims, the authors could also discuss its applicability to vision-language models (VLMs).
  3. W3. Mixed empirical evidence for the claimed improvements. While the authors emphasize the benefits of TopoU-Net for node classification, the empirical results do not consistently support a clear advantage over existing approaches. For instance, TopoU-Net outperforms the considered baselines on only approximately half of the node classification benchmarks. Furthermore, the comparison does not appear to include recent graph transformer architectures, which would be important baselines given the claims about modeling complex relational structures. The authors should better contextualize these mixed results and clarify under which conditions TopoU-Net provides meaningful improvements over existing methods.
Questions for authors
  1. Q1. Motivation for higher-order representations. In the abstract, the authors state that existing architectures often force structured data into grids, graphs, or sequences. Could the authors clarify why this is inherently problematic for the tasks considered in the paper? Given that a substantial portion of the evaluation focuses on node classification, graph classification, and hypergraph node classification, it is not immediately clear why standard graph-based representations would be insufficient. It would strengthen the motivation to provide concrete examples of structural information that is lost or cannot be adequately represented using conventional graph-based approaches.
  2. Q2. Expressiveness of GNNs and higher-order interactions. In the introduction, the authors suggest that GNNs are restricted to pairwise relational structures. However, this characterization seems somewhat broad, as several GNN variants and higher-order message-passing frameworks have been developed to model interactions beyond pairwise relationships. Could the authors clarify which specific limitations of existing GNN architectures motivate TopoU-Net? In particular, how does the proposed approach differ from or improve upon existing higher-order GNNs and hypergraph neural networks in terms of representational capacity and computational efficiency?
5 comments

We thank the reviewer for the careful evaluation and constructive feedback. We appreciate the recognition of the extensive experimental evaluation, the mathematical formulation, and the promising results on higher-order relational data, particularly hypergraphs. We agree that the motivation, relationship to existing higher-order GNNs, and conditions under which TopoU-Net is beneficial deserve further clarification. We address each concern below.

Weakness1. Clarity of Motivation and Contributions

We agree that the motivation and contributions can be communicated more directly.

The central contribution of TopoU-Net is not simply applying a U-Net to higher-order data, but introducing a general encoder--decoder formulation in which rank-wise cell representations define architectural levels, incidence operators provide cross-rank transport, and matched-rank skip connections preserve information during decoding.

Unlike conventional U-Nets, where the hierarchy is defined by spatial resolution, TopoU-Net organizes computation across the ranked relational structure of a combinatorial complex.

This provides a common architectural formulation for processing vertices, edges, and higher-order cells without requiring the same notion of spatial resolution across different domains.

The architectural design is supported by two complementary observations.

First, Section 3.4 introduces the bottleneck support ratio

ρbot=nsLns0\rho_{\mathrm{bot}} = \frac{n_{s_L}}{n_{s_0}}

which quantifies structural compression along a selected rank path. Proposition 3.6 further provides a necessary capacity condition for exact reconstruction through a linear bottleneck without skip connections.

Second, the controlled rank-path experiments in Appendix B.3.1 (Table 14) and the skip-connection ablations in Table 9 provide empirical evidence for the importance of higher-rank transport and matched-rank feature preservation.

We will revise the introduction to follow a clearer progression: (i) the motivation for explicit higher-order relational structure, (ii) rank-based encoder--decoder construction, and (iii) empirical evidence and limitations. We will also distinguish the architectural contribution from the supporting theoretical observations.

Weakness2. Comparisons with Modern Vision and Transformer Architectures

We appreciate this suggestion and agree that transformer-based comparisons would provide additional context.

However, the primary objective of TopoU-Net is to generalize the encoder--decoder principle to ranked relational domains, rather than to establish state-of-the-art performance against specialized vision architectures.

The image-segmentation experiments in Table 6 are intended to demonstrate the applicability of incidence-based encoder--decoder computation to spatial data through learned soft incidence operators.

We do not claim that these experiments establish superiority over modern vision transformers or vision-language models.

Similarly, graph transformers provide an important alternative for capturing long-range dependencies and global context. Their mechanisms, however, are different from the explicit incidence-based rank transport studied here.

We have not evaluated ViTs, graph transformers, or vision-language models directly and therefore make no claims of superiority over these architectures but We will work in this direction.

W3. Mixed Empirical Results and Conditions for Improvement

We agree that TopoU-Net does not uniformly outperform all baselines, and we do not claim universal superiority.

Our results instead suggest that the benefit of rank-wise encoding depends on the relational structure of the data and the information captured by the selected rank path.

For example, on Computers and Photo, GraphSAGE achieves stronger performance, indicating that conventional neighborhood aggregation can already be highly effective for these datasets.

In contrast, Table 4 reports the highest mean performance among the evaluated methods on four of five hypergraph benchmarks, where group-level interactions are explicitly represented.

Importantly, we investigate the contribution of higher-rank aggregation through controlled architectural ablations rather than relying only on comparisons with external baselines.

Capacity-controlled rank-path comparisons.

Appendix B.3.1 (Table 14) compares different rank paths under matched depth and comparable parameter budgets.

On Roman-empire, at 25% support, the edge-only path

0 → 1 → 0

achieves 68.56% accuracy, whereas

0 → 1 → 2 → 1 → 0

achieves 82.69%.

The corresponding parameter counts are 137,707 and 137,004, respectively.

At 50% support, the rank-2 path achieves 83.26% compared with 71.92% for the edge-only path.

These results demonstrate that higher-rank incidence aggregation can improve performance even without a global rank-3 cell, and that the observed differences are not explained solely by model size in these comparisons.

Contribution of skip connections.

Table 9 shows that removing skip connections reduces accuracy by 14.3, 14.5, and 12.1 percentage points on Texas, Wisconsin, and Cornell, respectively.

These datasets exhibit severe structural compression, consistent with the motivation for preserving matched-rank features during decoding.

Adaptive rank selection.

We additionally investigate learned soft rank routing, which adaptively weights admissible cross-rank transitions.

Compared with validation-selected fixed routing using the same routing backbone, learned routing improves Roman-empire accuracy from 86.39% to 88.32% and Amazon-ratings accuracy from 42.18% to 42.97%.

On Minesweeper, learned routing matches the best fixed-path accuracy of 83.98% while improving Macro-F1 from 71.76% to 72.91%.

These experiments suggest that the choice of rank path is consequential, and that adaptive selection can provide additional benefits.

Taken together, the results indicate that TopoU-Net is particularly promising when explicit higher-order relations provide useful information beyond local pairwise aggregation.

Nevertheless, the gains vary across datasets, and we will revise the experimental discussion to make these conditions and limitations more explicit.

Question 1. Motivation for Higher-Order Representations

We thank the reviewer for this important question. We agree that conventional graph representations are not inherently insufficient for every task, and that many graph-classification and node-classification problems can be addressed effectively by standard GNNs.

Furthermore, the combinatorial-complex formulation permits non-simplicial cells.

For example, in our MUTAG implementation, rank-2 cells are constructed using nx.cycle_basis(G), rather than being restricted to triangles.

Consequently, a cycle containing more than three vertices can be represented as a single rank-2 cell.

We additionally evaluate this principle on CSL using four-cycles as rank-2 cells. A four-cycle

c = {v₁, v₂, v₃, v₄}, rk(c) = 2,

is non-simplicial under this rank assignment, since a simplex containing four vertices has dimension three.

Using the rank path

0 → 1 → 2 → 1 → 0,

TopoU-Net achieves 18.67 ± 9.32% accuracy over 10 folds on CSL, compared with 11.33 ± 6.32% for GIN-pool under the same evaluation setup.

this experiment illustrates that TopoU-Net can operate on higher-order cells that are not simplices of the corresponding rank.

We emphasize that these representations are not the only way to preserve higher-order information. Incidence-based GNNs and other higher-order architectures can also model such structures.

Our contribution is therefore not the claim that graph neural networks are incapable of representing higher-order relations, but a systematic encoder--decoder organization for processing them across ranks.

We will clarify this distinction in the revised introduction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.