acceptodds
Under review as a conference paper at ICLR 2027

Topological Readouts of Learned Positional Filtrations for Hypergraph Classification

Abstract

Hypergraph neural networks perform hypergraph classification by message passing, which on graphs has been combined with positional encodings and persistent homology. Hypergraphs admit several expansions; a clique expansion can map hypergraphs whose star expansions differ in cycle rank to the same graph. On a widely used public benchmark, reported accuracies on two social collections lie near majority-class rates. Descriptors that combine positional and topological information on the star expansion may capture label information that these methods miss. We propose HyLPF: positional states initialized with random-walk positional encodings propagate over the star expansion, defining learned filtrations whose persistent homology, including cycle rank and essential-vertex flags, alone feeds the classifier. HyLPF achieves the best overall accuracy on the benchmark's four collections, ranking first on three and second on the fourth, with its configuration fixed before evaluation on new splits of the benchmark and under its split protocol and metric. On one of the social collections, matched controls show that the cycle rank and flags, with the weights reading them, carry label information, while merge values and cycle values add no detectable accuracy beyond them.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
hypergraph nerual network, position encoding, hypergraph learning

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_cQ1aOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a hypergraph classification method that constructs a star expansion from the incidence matrix and computes random-walk positional encodings (RWPE) for nodes and hyperedges using the diagonal entries of different powers of the random-walk transition matrices, representing return probabilities at different walk lengths. These encodings are propagated through message passing and used to define learned filtrations, whose persistent-homology features are aggregated for classification.

Critical strengths

Empirical improvements on two datasets: +8.88 percentage points on Steam-Player and +11.78 on Twitter-Friend over the strongest benchmark baselines on the same splits.

Critical weaknesses
  1. Limited novelty: The method mainly suggests positional encoding for hyper graphs, message passing, learned filtrations, and persistent-homology readouts. Its contribution is primarily their integration for hypergraph classification.

  2. Limited improvements across benchmarks: The two IMDB datasets show no statistically significant advantage over the strongest baselines. Moreover, approximately 28% (as mentioned in the manuscript) of their test graphs duplicate training or validation inputs.

Questions for authors

1.The methodological novelty appears limited. PiPE (Verma et al., ICML 2025) [1] already combines positional encodings, including RWPE, message passing, and persistent homology of learned filtrations. HyLPF applies closely related ideas to hypergraph star expansions with a different persistence readout. It remains unclear what additional representational capability this formulation provides over existing approaches. A comparison with PiPE adapted to the typed star expansion would help clarify the contribution.

[1] Positional Encoding meets Persistent Homology on Graphs Yogesh Verma, Amauri H Souza, Vikas K Garg

2 comments
AuthorsOpenReview ↗

Response to Reviewer cQ1a (1/2)

We thank the reviewer for the review and for the concrete request. Below, we address each point. All new experiments build on the code of the submission. They use the ten splits of the paper on Steam-Player and Twitter-Friend and three initializations per split. HyLPF was rerun together with each new model. We report every model we ran, and we will add all new results to the revised paper.


Q1 – Comparison with PiPE on the typed star expansion – We ran PiPE with its public code on the same star expansion and the same splits. HyLPF has the higher mean accuracy on both collections.

We added PiPE adapted to the typed star expansion, as you suggested.

  • Input: the star graph of HyLPF, with one node per vertex and per hyperedge, one edge per incidence, and a type flag in the node features.
  • Model: the layer classes of the public PiPE code, namely the GINE layers and the RePHINE persistence layer on learned filtrations of the random-walk encoding. An adapter connects them to our input features and to a classification head.
  • Configuration: depth, width, learning rate and dropout were selected from six configurations by validation accuracy on three splits that are not used for evaluation.

Table R3: PiPE on the typed star expansion. Test accuracy in % (mean ± SD over 10 splits) and paired difference HyLPF − PiPE. Every difference has a 95% interval that excludes zero. As in the paper, † marks a difference whose overlap-corrected interval (Appendix C.1) contains zero. Bold marks the best value in each column.

ModelSteam-PlayerTwitter-FriendDiff. Steam-PlayerDiff. Twitter-Friend
HyLPF66.50 ± 2.6573.05 ± 2.28——
PiPE, public code, on the typed star expansion64.37 ± 2.1468.58 ± 3.20+2.14†+4.48

As shown in Table R3, HyLPF has the higher mean accuracy on both collections, by 2.14 points on Steam-Player and by 4.48 points on Twitter-Friend. Both differences have 95% intervals that exclude zero, and on Twitter-Friend the overlap-corrected interval also excludes zero. In macro-F1, HyLPF is 2.33 points higher on Steam-Player and 6.44 points higher on Twitter-Friend. Both intervals again exclude zero on Twitter-Friend. In this run HyLPF reaches 66.50% and 73.05%, in agreement with the paper (66.41% and 72.93%).

PiPE on the star expansion is also above the 18 benchmark methods of Table 1 in mean accuracy. Their best values are 57.52% on Steam-Player and 61.15% on Twitter-Friend. This supports the typed star graph with random-walk encodings as a representation for these collections.

In the revision, we will add Table R3 and list every adaptation of the PiPE code.

AuthorsOpenReview ↗

Response to Reviewer cQ1a (2/2)

Q1 – Additional capability over PiPE – HyLPF differs from PiPE in four respects, and a test that changes one setting shows that the main difference matters.

Both methods combine positional encodings, message passing and persistence of learned filtrations. The paper cites PiPE in Section 2. HyLPF differs in four respects.

  1. Domain. HyLPF is defined on hypergraphs through the typed star graph, with separate filtration networks for nodes and hyperedges. Proposition 6 gives two hypergraphs with the same clique graph and the same degrees whose star graphs have cycle ranks 3 and 5. The readout receives the cycle rank, so it separates them.
  2. Source of the filtrations and input of the classifier. PiPE predicts from pooled node embeddings together with a topological embedding. In HyLPF, the positional states are propagated separately from the features, and they alone define the filtrations. The classifier reads only the persistence descriptor.
  3. Ties. Vertices exchanged by an automorphism receive equal filtration values, so ties are systematic. Definition 3 and Propositions 4 and 5 give a descriptor that is well defined and isomorphism invariant under ties.
  4. Evaluation. Two matched twins separate the persistence-derived inputs from the persistence values (Definition 7, Table 2).

Relative to PiPE on the same star graph, the difference lies in point 2. We tested this difference directly in our own framework, on the same splits. The variants below ran on a different machine from PiPE, so Table R2 has its own rerun of HyLPF.

Table R2: Variants on the typed star expansion. Test accuracy in % (mean ± SD over 10 splits) and paired difference HyLPF − model. Every difference excludes zero under the 95% interval and under the overlap-corrected interval (Appendix C.1). Bold marks the best value in each column.

ModelSteam-PlayerTwitter-FriendDiff. Steam-PlayerDiff. Twitter-Friend
HyLPF66.28 ± 1.8772.93 ± 2.31——
One switch changed: filtrations from the feature states62.99 ± 3.4465.11 ± 3.22+3.28+7.81
GFL-style, implemented in our framework58.85 ± 1.9264.50 ± 2.28+7.43+8.42
TOGL-style, implemented in our framework57.30 ± 2.9762.85 ± 4.74+8.98+10.08

As shown in Table R2, the first variant changes one setting so that the filtrations come from the feature states instead of the positional states. This lowers the accuracy by 3.28 points on Steam-Player and by 7.81 points on Twitter-Friend. Two further variants follow the designs of GFL (Hofer et al., 2020) and TOGL (Horn et al., 2022). Their accuracies are 7.4 to 10.1 points lower. All six differences exclude zero under both intervals. The three variants use our tie-aware descriptor and the configuration of HyLPF.

In the revision, we will add Table R2 and a paragraph in Section 2 that states the differences from PiPE item by item.


W1 – Limited novelty – The components exist separately. HyLPF combines them in a new design with a tie-aware descriptor, and the new comparisons show that this design matters.

We agree that positional encodings, message passing, learned filtrations and persistence readouts exist separately. The four points above describe what is new. The tie-aware descriptor in point 3 is a technical contribution, as earlier work on learned filtrations does not address ties. Table R2 shows that the design choices of HyLPF raise the accuracy on the same representation. HyLPF also has a higher mean accuracy than PiPE in Table R3.


W2 – IMDB results and duplicated test graphs – We agree. The paper claims no significant difference on the IMDB collections, and the duplicates were found by our own audit.

The paper reports mean accuracies on the two IMDB collections, where HyLPF is first on IMDB-Dir-Genre and second on IMDB-Dir-Form. Every contrast whose interval excludes zero is on a social collection. Appendix D reports the duplicates so that readers can weigh the IMDB numbers accordingly.


We hope the comparison with PiPE and these answers resolve your concerns. If they do, we would be grateful if you would reconsider your score and confidence.

Reviewer_MTqmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

HyLPF learns filtrations on a hypergraph's typed star expansion and classifies the hypergraph using persistence-derived merge values, cycle information, and essential flags. The strongest reported gains are on Steam-Player and Twitter-Friend.

Critical strengths

The Steam-Player result holds up: HyLPF beats the paper's own 22-feature count regression by 6.65 points, and the corrected interval [3.52, 9.78] stays above zero (Table 12). The matched controls and paired comparisons are carefully reported. The duplicate-graph audit is also helpful to anyone using the IMDB benchmarks.

Critical weaknesses
  1. In the value-ablated twin (Definition 7, Section 3.4), zero-dimensional merge values are replaced by the underlying filtration values, and one-dimensional cycle values by the constant 1/2. The essential flags and cycle-token counts stay. Full HyLPF gains only +0.02 points on Steam and +0.31 on Twitter relative to that twin, with both intervals spanning zero (Tables 2 and 10).

    For Steam, the authors themselves conclude that persistence merge/cycle values add no detectable accuracy beyond cycle rank and essential flags. That leaves me wondering what the extra persistence-value computation buys. Cycle rank is already among the 22 count-regression features (Appendix C).

  2. On both social datasets, the count regression beats the reported mean accuracy of all 18 original benchmark methods. HyLPF still wins clearly on Steam, but the +4.38-point advantage on Twitter has a corrected interval of [-0.66, 9.42] (Table 12). A GIN or GINE on the same star graph with RWPE, counts, and flags would be a more revealing matched baseline.

  3. The configuration was selected through cross-validation over the same graph collections that were later re-split for evaluation (Appendix B). Consequently, those later splits are not fully independent of model selection, even though the procedure is disclosed.

  4. About 28% of IMDB test graphs have isomorphic copies in training or validation: 28.5% for IMDB-Dir-Form and 27.7% for IMDB-Dir-Genre (Appendix D). The IMDB ablation contrasts are inconclusive as well, so I would not put much weight on small differences there.

Minor Comment

The number of one-dimensional cycle tokens is the star graph's cycle rank (Proposition 6). For the incidence graph, the standard identity is β1\beta_1 = |I| − |V| − |E| + β0\beta_0. This closed-form count could be distinguished more clearly from computing numerical persistence values.

Questions for authors

On truly independent data, does the full persistence-value readout beat a parameter-matched star-graph GNN that uses RWPE, cycle counts, and essential flags?

2 comments
AuthorsOpenReview ↗

Response to Reviewer MTqm (1/2)

We thank the reviewer for the precise reading of Tables 2, 10 and 12 and of Definition 7, and for recognizing the Steam-Player result, the matched controls and the duplicate audit. Below, we address each point. All new experiments build on the code of the submission. They use the ten splits of the paper on Steam-Player and Twitter-Friend and three initializations per split. HyLPF was rerun together with the new models and reached 66.28% and 72.93% (paper: 66.41% and 72.93%). We report every model we ran, and we will add all new results to the revised paper.


W2, Q1 – Matched star-graph baseline with RWPE, counts and flags – We ran this baseline in three versions. HyLPF has higher mean accuracy than each of them on both collections.

On the same typed star graph, with RWPE and the configuration of HyLPF, we added three baselines that compute no persistence.

  • B1: the message-passing network of HyLPF (Equations 2 and 6). The classifier reads the sum-pooled node and hyperedge states.
  • B2: B1 with the 22 count features of Appendix C, which include the cycle rank, as additional classifier inputs.
  • B3: B2 with the closest analogue of the essential flags that needs no filtration. By Proposition 4, a flag marks the minimum of a filtration on a connected component. B3 therefore pools the minimum of the final states over each component.

Table R1: Test accuracy in % (mean ± SD over 10 splits) and paired difference HyLPF − model. Every difference has a 95% interval that excludes zero. As in the paper, † marks a difference whose overlap-corrected interval (Appendix C.1) contains zero. Bold marks the best value in each column.

ModelSteam-PlayerTwitter-FriendDiff. Steam-PlayerDiff. Twitter-Friend
HyLPF66.28 ± 1.8772.93 ± 2.31——
B1: same backbone, no persistence64.06 ± 3.0568.30 ± 2.76+2.22†+4.63
B2: B1 + 22 graph invariants63.62 ± 2.4869.77 ± 4.45+2.65+3.16†
B3: B2 + component-wise minimum pooling63.77 ± 2.3068.88 ± 3.41+2.51+4.05

As shown in Table R1, HyLPF has the highest mean accuracy on both collections, 2.2 to 4.6 points above the three baselines. All six differences have 95% intervals that exclude zero, and four of them also do so under the overlap correction. In macro-F1, HyLPF is also higher in all six comparisons, by 2.2 to 2.7 points on Steam-Player and by 4.9 to 7.1 points on Twitter-Friend.

Three further facts place this result in context.

  1. Parameters. The baselines have 634k to 661k parameters on Steam-Player. HyLPF has 738k, of which 231k enter its prediction (Section 5.2).
  2. GINE backbone. PiPE on the same star graph uses GINE layers with RWPE and a persistence branch, and has its own configuration search. It reaches 64.37% and 68.58%, and HyLPF reaches 66.50% and 73.05% in the same run (Table R3 in our response to Reviewer cQ1a).
  3. Benchmark methods. B1 alone reaches 64.06% and 68.30%. The best of the 18 benchmark methods reaches 57.52% and 61.15% on the same splits (Table 1). The star-graph backbone with propagated RWPE is therefore effective on its own, and HyLPF is 2.2 and 4.6 points above it.

As you note, the corrected interval for the count regression on Twitter-Friend contains zero (Table 12). The new baselines provide the more revealing comparison that you asked for. Regarding independent data, every collection in the benchmark took part in the configuration search, so we report results on the ten splits of the paper. Our reply to W3 below measures how much the result depends on the selected configuration.

In the revision, we will add Table R1 to Section 5.2 and Appendix C.


W1 – What the persistence values buy – On Steam-Player, the gain comes from the essential flags and the cycle rank. The numerical values add no detectable accuracy, as the paper reports.

We agree with your reading of Tables 2 and 10. The readout has two parts, and we will describe them separately in Sections 3.3 and 5.2.

  • Topological summary: the essential flags and the number of cycle tokens. On Steam-Player, this part carries the gain. The value-ablated twin exceeds the homology-matched twin by 3.69 points, and both intervals exclude zero (Table 13).
  • Numerical values: the merge values and cycle values. They add 0.02 points on Steam-Player and 0.31 points on Twitter-Friend, and both intervals contain zero.

Both parts come from one union-find pass over the sorted incidences. The summary alone needs only the connected components and the minimum of each filtration on each component.

The cycle rank is among the 22 count features, as you say. HyLPF still gives more than these counts. On Steam-Player, it exceeds the count regression by 6.65 points (Table 12) and B2 by 2.65 points (Table R1), and both intervals exclude zero in each case.

AuthorsOpenReview ↗

Response to Reviewer MTqm (2/2)

W3 – Configuration selected on the same collections – We agree, and we now measure this dependence. Six of eight configurations drawn at random from the search space also exceed the benchmark methods on both collections.

The statement in Section 4 stays in the paper. We added a study with eight other configurations, drawn at random from the search space of Appendix B before we ran any of them. Each was trained on the same ten splits with three initializations. No selection step is involved, and all eight are reported.

Table R4: Test accuracy in % (mean ± SD over 10 splits) of the selected configuration and of eight configurations drawn at random from the search space. The last row is the best of the 18 benchmark methods (Table 1). Bold marks the best value in each column.

ConfigurationClassifier inputSteam-PlayerTwitter-Friend
Selected (HyLPF)readout66.28 ± 1.8772.93 ± 2.31
Draw 1readout and pooled states55.42 ± 3.9562.34 ± 6.91
Draw 2readout and pooled states63.98 ± 2.9070.05 ± 2.43
Draw 3readout and pooled states64.05 ± 3.4271.30 ± 4.34
Draw 4readout and pooled states64.82 ± 2.3770.03 ± 2.95
Draw 5readout and pooled states62.06 ± 3.6168.50 ± 2.85
Draw 6readout64.77 ± 2.4671.96 ± 2.58
Draw 7readout and pooled states64.84 ± 2.2869.16 ± 2.72
Draw 8readout54.00 ± 3.5963.36 ± 4.42
Best benchmark method—57.5261.15

As shown in Table R4, six of the eight random configurations reach 62.1% to 64.8% on Steam-Player and 68.5% to 72.0% on Twitter-Friend. The best benchmark method reaches 57.52% and 61.15%. The medians of the eight are 64.01% and 69.59%. This accuracy level relative to the benchmark methods is therefore not specific to the single selected configuration. The selected configuration has the highest mean of the nine, so the statement in Section 4 remains unchanged.

In the revision, we will add Table R4 to Appendix B.


W4 – IMDB duplicates and IMDB contrasts – We agree, and the paper claims no significant difference on the IMDB collections.

Thank you for finding the audit helpful. The paper reports mean accuracies on the two IMDB collections, where HyLPF is first on IMDB-Dir-Genre and second on IMDB-Dir-Form. Every contrast whose interval excludes zero is on a social collection. Appendix D reports the duplicates so that readers can weigh the IMDB numbers accordingly.


Minor comment – Closed form of the cycle rank – We agree, and we will state the identity next to Proposition 6.

We will add β1(S(H))=∣I∣−∣V∣−∣E∣+β0(S(H))\beta_1(S(H)) = |I| - |V| - |E| + \beta_0(S(H)) after Proposition 6. We will also say that the number of cycle tokens follows from this identity without any persistence computation.


We hope the matched baselines and these clarifications resolve your concerns. If they do, we would be grateful if you would reconsider your score and confidence.

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

The paper proposes HyLPF, a hypergraph classification method that uses random-walk positional encodings and message passing on the star expansion of a hypergraph to learn vertex filtrations. The persistent homology of these filtrations is then aggregated through a DeepSets-style readout for classification.

The authors provide theoretical results concerning their persistence descriptor, isomorphism invariance, and differences between clique and star expansions. The method is evaluated on four hypergraph classification datasets, reporting improvements over several existing hypergraph neural networks.

Critical strengths

I do think that the paper is strong in the range of experiments. Also their solve for tie in filtration functions is interesting and something that some of the previous works do not discuss. So although there are some missing baselines/better discussion with the current literature missing in the experiments, I do think that the authors made a good effort in experiments.

Critical weaknesses

My main concern is that most of the proposed mechanism has already been introduced in previous works. Graph Filtration Learning [1] already introduced learnable vertex filtrations and differentiable persistent homology as a readout for graph classification, which is essentially the central idea behind HyLPF. Similarly, Topological Graph Neural Networks [2] explored persistent homology from learned graph representations, while Positional Encoding meets Persistent Homology on Graphs [3] specifically investigated combining positional encodings with persistent homology. PH-based hypergraph classification has also been studied previously [4].

Other relevant works missing from the discussion include Persistent Weisfeiler-Lehman for graph classification [5], filtration-based structural descriptors [6], the expressivity of PH in graph learning [7], and learnable topological positional encodings [8].

From what I understood, HyLPF mainly combines these existing ideas on the star expansion of a hypergraph. Moreover, this expansion is fixed, and the authors do not learn new higher-order structures. This is particularly limiting considering that learnable higher-order constructions have already been explored by DCM [9] and DiffLift [10]. Comparisons with these approaches, appropriately adapted to this setting, would help evaluate the choice of a fixed representation.

Another point is that I couldnt understand what we are gaining from using PH here, if removing the persistence values does not produce a detectable accuracy reduction, what is actually gained from computing learned persistent homology?

Related work showing the effectiveness of simple graph invariants [11] further motivates comparisons against simpler structural baselines.

Therefore, I think stronger comparisons with simpler structural readouts and existing learned-filtration methods are necessary to establish whether the proposed PH mechanism is responsible for the reported improvements.


[1] C. Hofer et al. Graph Filtration Learning. ICML, 2020. https://proceedings.mlr.press/v119/hofer20b.html

[2] M. Horn et al. Topological Graph Neural Networks. ICLR, 2022. https://openreview.net/forum?id=oxxUMeFwEHd

[3] Y. Verma, A. H. Souza, and V. K. Garg. Positional Encoding meets Persistent Homology on Graphs. ICML, 2025. https://proceedings.mlr.press/v267/verma25b.html

[4] M. E. Aktas et al. Persistent Homology of Hypergraphs via Barycentric Subdivision. Journal of Complex Networks, 2026. https://doi.org/10.1093/comnet/cnag013

[5] B. Rieck, C. Bock, and K. Borgwardt. A Persistent Weisfeiler-Lehman Procedure for Graph Classification. ICML, 2019. https://proceedings.mlr.press/v97/rieck19a.html

[6] L. O'Bray, B. Rieck, and K. Borgwardt. Filtration Curves for Graph Representation. KDD, 2021. https://doi.org/10.1145/3447548.3467442

[7] R. Ballester and B. Rieck. On the Expressivity of Persistent Homology in Graph Learning. LoG, 2025. https://proceedings.mlr.press/v269/ballester25a.html

[8] J. P. Amboage, E. Röell, P. Schnider, and B. Rieck. LEAP: Local ECT-Based Learnable Positional Encodings for Graphs. ICLR, 2026. https://proceedings.iclr.cc/paper_files/paper/2026/hash/45c66714da9f050814df6bc056ebe8a3-Abstract-Conference.html

[9] C. Battiloro et al. From Latent Graph to Latent Topology Inference: Differentiable Cell Complex Module. ICLR, 2024. https://proceedings.iclr.cc/paper_files/paper/2024/hash/6b97236d90d945be7c58268207a14f4f-Abstract-Conference.html

[10] J. Franco et al. Differentiable Lifting for Topological Neural Networks. ICLR, 2026. https://proceedings.iclr.cc/paper_files/paper/2026/hash/49691197eef302e1343a37229aa947d9-Abstract-Conference.html

[11] R. von Moos, M. Alain, and B. Rieck. Measuring Structure in Graph Benchmark Datasets Using Graph Invariants. arXiv preprint, 2026. https://arxiv.org/abs/2605.06462

Questions for authors
  1. What exactly is the main methodological novelty of HyLPF compared with Graph Filtration Learning [1], TOGL [2], and PiPE [3]? Since the method applies learned filtrations and persistent homology to a fixed star expansion, what fundamentally distinguishes it from applying these existing approaches to the same representation? Also, why restrict the topological domain to a fixed one-dimensional graph and not considering static/learnable higher-order liftings?

  2. Given the results in Table 2, where removing the persistence values does not seem produce a reduction in accuracy, what is actually gained from computing learned persistent homology? Could the authors compare against a simpler model using the same configs without computing persistence diagrams?

5 comments
AuthorsOpenReview ↗

Response to Reviewer Ycwv (1/3)

We thank the reviewer for the detailed review, for the eleven references, and for noting our treatment of ties in the filtration functions. Below, we address each point. References [1] to [11] are those of the review. All new experiments build on the code of the submission. They use the ten splits of the paper on Steam-Player and Twitter-Friend and three initializations per split. HyLPF was rerun together with the new models and reaches 66.28% and 72.93% (paper: 66.41% and 72.93%). We report every model we ran, and we will add all new results to the revised paper.


Q2, W4, W5 – Gain from persistent homology and simpler baselines – We ran the model you asked for. HyLPF has higher mean accuracy on both collections, even when the simpler model receives graph invariants.

We added three simpler models that keep the backbone and the configuration of HyLPF and compute no persistence diagram.

  • B1: the message passing of HyLPF on the typed star graph with RWPE. The classifier reads the sum-pooled node and hyperedge states. This is the model in Q2.
  • B2: B1 with the 22 graph invariants of Appendix C as additional classifier inputs. They include the number of components and the cycle rank, in the spirit of [11].
  • B3: B2 with pooling that needs connected components only, namely the component-wise minimum of the final states.

Table R1: Test accuracy in % (mean ± SD over 10 splits) and paired difference HyLPF − model. Every difference has a 95% interval that excludes zero. As in the paper, † marks a difference whose overlap-corrected interval (Appendix C.1) contains zero. Bold marks the best value in each column.

ModelSteam-PlayerTwitter-FriendDiff. Steam-PlayerDiff. Twitter-Friend
HyLPF66.28 ± 1.8772.93 ± 2.31——
B1: same backbone, no persistence64.06 ± 3.0568.30 ± 2.76+2.22†+4.63
B2: B1 + 22 graph invariants63.62 ± 2.4869.77 ± 4.45+2.65+3.16†
B3: B2 + component-wise minimum pooling63.77 ± 2.3068.88 ± 3.41+2.51+4.05

As shown in Table R1, HyLPF has the highest mean accuracy on both collections, 2.2 to 4.6 points above the three simpler models. All six differences have 95% intervals that exclude zero, and four also have overlap-corrected intervals that exclude zero. The control in the paper agrees. The homology-matched twin keeps the architecture and withholds every persistence-derived input. It is 3.71 points lower on Steam-Player, and both intervals exclude zero (Table 2).

What is gained. On Steam-Player, where the twins in the paper isolate this effect, the gain comes from the topological summary that the readout extracts from the learned filtrations. It does not come from the numerical persistence values.

  1. The summary consists of the essential flags and the cycle tokens. A flag marks where a learned filtration attains its minimum on a connected component (Proposition 4). The number of cycle tokens is the cycle rank of the star graph (Proposition 6). On Steam-Player, this summary carries the gain: the value-ablated twin exceeds the homology-matched twin by 3.69 points, and both intervals exclude zero (Table 13).
  2. The merge values and cycle values add no detectable accuracy, as Table 2 reports. We do not claim otherwise.
  3. Table R1 adds one new fact. With the cycle rank, the other invariants, and the component-wise minima of its states, the same backbone (B2 and B3) stays below HyLPF in mean accuracy on both collections.

Regarding [11], the paper already has a baseline of this kind: a logistic regression on the 22 invariants (Table 12). HyLPF exceeds it by 6.65 points on Steam-Player, where both intervals exclude zero, and by 4.38 points in mean accuracy on Twitter-Friend. B2 is a stronger version of this baseline.

B1 alone reaches 64.06% and 68.30%. The best of the 18 benchmark methods reaches 57.52% and 61.15% on the same splits (Table 1). The typed star graph with propagated RWPE is therefore an effective representation on its own, and HyLPF is 2.2 and 4.6 points above B1.

In the revision, we will add Table R1 and describe the two parts of the readout separately in Sections 3.3 and 5.2.

AuthorsOpenReview ↗

Response to Reviewer Ycwv (2/3)

Q1 (first part), W1 – Novelty over GFL [1], TOGL [2] and PiPE [3] – HyLPF differs from them in four respects, and we now compare with all three designs on the same star expansion.

We agree that learned filtrations with a differentiable persistence readout originate in [1], [2] and [3]. The paper cites them in Section 2, together with [4], which uses hand-designed filtrations and does not learn them. HyLPF differs in four respects.

  1. Domain. The filtrations live on the typed star graph of a hypergraph, with separate filtration networks for nodes and hyperedges. Proposition 6 gives two hypergraphs with the same clique graph and the same degrees whose star graphs have cycle ranks 3 and 5. The readout receives the cycle rank.
  2. Source of the filtrations and input of the classifier. In GFL, a network on node features defines the filtration. TOGL adds persistence features back to the node states. PiPE predicts from pooled node embeddings together with a topological embedding. In HyLPF, the positional states are propagated separately from the features, and they alone define the filtrations. The classifier reads only the persistence descriptor.
  3. Ties. Vertices exchanged by an automorphism receive equal filtration values, so ties are systematic. Definition 3 and Propositions 4 and 5 give a descriptor that is well defined and isomorphism invariant under ties.
  4. Evaluation. Two matched twins separate the persistence-derived inputs from the persistence values (Definition 7).

We tested point 2 on the same typed star expansion, with three variants (Table R2) and with the public PiPE code (Table R3).

Table R2: Variants on the typed star expansion. Test accuracy in % (mean ± SD over 10 splits) and paired difference HyLPF − model. Every difference excludes zero under the 95% interval and under the overlap-corrected interval (Appendix C.1). Bold marks the best value in each column.

ModelSteam-PlayerTwitter-FriendDiff. Steam-PlayerDiff. Twitter-Friend
HyLPF66.28 ± 1.8772.93 ± 2.31——
One switch changed: filtrations from the feature states62.99 ± 3.4465.11 ± 3.22+3.28+7.81
GFL-style, implemented in our framework58.85 ± 1.9264.50 ± 2.28+7.43+8.42
TOGL-style, implemented in our framework57.30 ± 2.9762.85 ± 4.74+8.98+10.08

As shown in Table R2, changing one setting so that the filtrations come from the feature states instead of the positional states lowers the accuracy. Accuracy drops by 3.28 points on Steam-Player and by 7.81 points on Twitter-Friend. The GFL-style and TOGL-style variants are 7.4 to 10.1 points lower. All six differences exclude zero under both intervals. The three variants implement these designs in our framework, with our tie-aware descriptor and the configuration of HyLPF.

For Table R3, we built PiPE on the typed star graph from the layer classes in its public code. These are the RePHINE persistence layer and the GINE layers of the random-walk branch. Its depth, width, learning rate and dropout were selected from six configurations on splits that are not used for evaluation. PiPE ran on a second machine, so Table R3 has its own rerun of HyLPF.

Table R3: PiPE on the typed star expansion. Test accuracy in % (mean ± SD over 10 splits) and paired difference HyLPF − PiPE. Every difference has a 95% interval that excludes zero. As in the paper, † marks a difference whose overlap-corrected interval (Appendix C.1) contains zero. Bold marks the best value in each column.

ModelSteam-PlayerTwitter-FriendDiff. Steam-PlayerDiff. Twitter-Friend
HyLPF66.50 ± 2.6573.05 ± 2.28——
PiPE, public code, on the typed star expansion64.37 ± 2.1468.58 ± 3.20+2.14†+4.48

HyLPF has a higher mean accuracy than PiPE on both collections. Both differences have 95% intervals that exclude zero, and the difference on Twitter-Friend also excludes zero under the overlap correction.

In the revision, we will add Tables R2 and R3 with a description of every model.

AuthorsOpenReview ↗

Response to Reviewer Ycwv (3/3)

Q1 (second part), W3 – Fixed star expansion and learnable liftings [9, 10] – The hypergraph is the given input, and the typed star graph encodes it exactly, so no higher-order structure has to be inferred.

The typed star graph of Definition 1 has the nodes, hyperedges and incidences of the input hypergraph. Nothing is lost, and nothing has to be learned to obtain the higher-order structure.

DCM [9] and DiffLift [10] address the complementary case. They infer cells or hyperedges from a point cloud or a graph that carries no higher-order structure. Our inputs already carry this structure. A comparison would require us to discard the given hyperedges and infer them again. Neither paper specifies this for hypergraph inputs, so we did not run these methods. A learned lifting on top of the given incidences is a natural extension of our work.

The one-dimensional domain is chosen for cost and fidelity. Persistence in dimensions 0 and 1 on the star graph costs O(∣I∣log⁡∣I∣)O(|I| \log |I|) per filtration. A simplicial closure of the hyperedges would add faces that are not in the data. Proposition 6 shows that the star graph can separate hypergraphs that share one clique graph.

We will add this discussion and both references to Section 2.


W2 – Related work [5] to [8] and [11] – Thank you. We will discuss all of them in the paper.

We read all five works. We will add them as follows.

  • Persistent Weisfeiler-Lehman [5] and filtration curves [6]: Section 2, as descriptors built on fixed filtrations. HyLPF learns its filtrations.
  • Expressivity of persistent homology [7]: Section 2.
  • LEAP [8]: Section 2, as a learnable topological positional encoding.
  • Graph invariants [11]: Section 5.2, where the count baseline of Table 12 is introduced, together with B2 of Table R1.

We hope these answers and the new comparisons resolve your concerns. If they do, we would be grateful if you would reconsider your score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

To the Area Chair and all Reviewers: summary of the discussion and the revision. Thank you!

Dear Area Chair and Reviewers,

We sincerely thank the area chair and all three reviewers for the time and care given to our submission. The discussion has made the paper stronger. As the discussion closes, we summarize what the paper contributes, what was asked and added, and our view on the remaining concern.


Contributions – What the paper offers to the hypergraph learning and topological deep learning communities.

  • A new mechanism. To our knowledge, HyLPF is the first method that reads the persistent homology of learned positional filtrations on the star expansion of a hypergraph. The classifier reads only this descriptor.
  • A tie-aware descriptor with proofs. Learned filtration values tie systematically under automorphisms. Our descriptor is well defined and isomorphism invariant under ties. Reviewer Ycwv notes that this is "something that some of the previous works do not discuss".
  • Results under a public protocol. HyLPF achieves the best overall accuracy on the four collections of the benchmark. On the two social collections, it exceeds the best of 18 benchmark methods by 8.88 and 11.78 points on the same splits. Both intervals exclude zero.
  • Careful evaluation that others can reuse. The configuration was frozen before evaluation. Matched twins separate what the readout uses. An audit shows that about 28% of the IMDB test graphs have isomorphic copies in training or validation. Reviewer MTqm calls the audit "helpful to anyone using the IMDB benchmarks".

Discussion – We ran every experiment that the reviewers requested, and we report every model.

  • Ycwv and MTqm ask for simpler models without persistence: added (Table R1). HyLPF has the highest mean accuracy on both collections, 2.2 to 4.6 points above three such models. All six 95% intervals exclude zero.
  • Ycwv and cQ1a ask for GFL, TOGL and PiPE on the same star expansion: added (Tables R2 and R3). Changing one setting, so that the filtrations come from the feature states, lowers the accuracy by 3.28 and 7.81 points. GFL-style and TOGL-style variants are 7.4 to 10.1 points lower. HyLPF also has a higher mean accuracy than PiPE built from its public code, by 2.14 and 4.48 points.
  • MTqm asks how much the result depends on the selected configuration: measured (Table R4). Six of eight configurations drawn at random reach 62.1% to 64.8% and 68.5% to 72.0%. The best benchmark method reaches 57.52% and 61.15%.
  • Ycwv and MTqm ask for related work and the closed form of the cycle rank: added.

The revised paper contains all of this in blue. Every number in the submitted version is unchanged. The new tables keep the names R1 to R5 used in our responses.


Remaining concern – Our view on the necessity of persistent homology and on learnable liftings.

We thank Reviewer Ycwv for reading our responses. We respect the two remaining concerns and offer our view.

  1. Persistence values. The paper itself shows with a matched control that the numerical persistence values add no detectable accuracy, and it says so in the abstract. We see this as a finding. It tells the community which part of the readout carries the signal. On Steam-Player, the topological summary of the learned filtrations, namely the essential flags and the cycle rank, gives a gain of 3.69 points over a matched control without it. When the same backbone receives the cycle rank and component-wise minima directly, it stays 2.5 to 2.7 points below HyLPF there. Both intervals exclude zero in each case.
  2. Learnable liftings. Our inputs are hypergraphs, so the higher-order structure is given, and the typed star graph encodes it exactly. Learned liftings address inputs without such structure. We agree that combining the two is a promising direction.

All three reviews rate the soundness concerns as minor or fixable, and we have answered each of them with new experiments. We believe that the paper offers a new mechanism, proved properties, and carefully controlled evidence, and that the discussion has strengthened each of them. We are grateful to everyone for this constructive process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.