Pith. sign in

REVIEW 4 major objections 6 minor 30 references

Topo-Scan turns graphs into short topological token sequences that Transformers can read, matching or beating strong GNN and homology baselines without full persistence diagrams.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-31 13:02 UTC pith:H2YZF62M

load-bearing objection Solid methods paper: sliding-window interlevel Betti tokens for Transformers beat same-filtration PH baselines and look competitive, with real runtime wins; SOTA claims need tempering on small-data variance and hyperparameter selection. the 4 major comments →

arxiv 2607.28259 v1 pith:H2YZF62M submitted 2026-07-30 cs.LG math.AT

TopoFormer: Topology Meets Attention for Graph Learning

Md Joshem Uddin , Astrit Tola , Cuneyt Gurcan Akcora , Baris Coskunuzer This is my paper
classification cs.LG math.AT
keywords graph representation learningTopo-Scanpersistent homologyTransformerstopological data analysisgraph classificationmolecular property predictioninterlevel Betti numbers
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

TopoFormer claims that multi-scale graph structure can be fed to ordinary Transformers by first slicing a graph along a filtration into overlapping windows and recording a few simple numbers per window—connected components, cycles, node count, and edge count. That module, Topo-Scan, deliberately drops the nested filtration and expensive persistence-diagram step of classical topological data analysis, keeping only ordered, parallelizable slice tokens. The paper argues these tokens stay stable when the filtration signal is perturbed, capture late-emerging structure that ordinary sublevel filtrations often bury, and, once attended over, produce graph-level embeddings competitive with strong graph neural nets and topology baselines on classification and molecular property tasks. A sympathetic reader cares because the method offers a fixed-budget, attention-native way to inject topological inductive bias without the usual computational bottlenecks or bespoke vectorizations.

Core claim

A short ordered sequence of interlevel Betti numbers and size counts, obtained by sliding a window over a node or edge filtration on the clique complex, is a stable, Transformer-ready summary of multi-scale graph topology that matches or exceeds strong GNN and persistent-homology baselines while remaining fully parallelizable and free of persistence-diagram computation.

What carries the argument

Topo-Scan: slice a graph into overlapping induced subgraphs along thresholds of a filtration function, compute (β0, β1, |V|, |E|) on each slice’s clique complex, and feed the resulting fixed-length token sequence to a Transformer; stability follows from bottleneck control of the underlying interlevel persistence modules.

Load-bearing premise

The four simple numbers per slice, on a fixed clique skeleton with a small hand-chosen window width and a few fixed filtrations, keep enough task-relevant structure that the gains come from the sequential Topo-Scan idea itself rather than from fusion tricks or model capacity alone.

What would settle it

On the same filtration, thresholds, and clique 2-skeleton, replace Topo-Scan windows by ordinary sublevel Betti vectors (or randomize slice order) and check whether the Transformer’s accuracy and late-threshold Betti activity collapse to the weaker PH-MLP/PH-TR baselines reported in the ablations.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Graph-level models can inject multi-scale topology as short token sequences without global boundary-matrix reductions or custom persistence vectorizations.
  • Late-emerging components and cycles that classical sublevel filtrations bury remain visible to attention because slices are range-localized rather than cumulative.
  • Stability of the discrete token sequences is controlled by the bottleneck distance of interlevel modules, hence by the sup-norm of the filtration signal.
  • The same backbone can fuse multiple filtrations by independent Transformers plus learnable attention, and can be hybridized with molecular fingerprints for property prediction.
  • Fixed-length topological sequences are a natural substrate for future topology-aware graph foundation-model pretraining.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If Topo-Scan tokens transfer across domains, pretraining a single Transformer on heterogeneous graph corpora of these sequences could yield a lightweight topology-only foundation model that never sees raw nodes.
  • Learnable or task-conditioned filtrations plugged into the same sliding-window backbone would test whether the method’s gains survive when the filtration is no longer hand-chosen.
  • Extending per-slice tokens beyond (β0, β1, sizes) to richer local invariants, while keeping sequence length fixed, is a direct route to higher expressivity without restoring full persistence diagrams.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes TopoFormer, a graph-level architecture whose core module, Topo-Scan, replaces nested sublevel persistent homology with overlapping interlevel slices of a fixed clique complex under scalar node/edge filtrations. Each slice is summarized by a four-dimensional token (β0, β1, |Vi|, |Ei|); the resulting short ordered sequences are fed to Transformer encoders (optionally fused across filtrations or with molecular fingerprints as TopoFormer*). The authors prove discrete ℓ1 stability of the Topo-Scan sequences with respect to the bottleneck distance of interlevel modules (Theorem 3.1, Corollary 3.2, Appendix B), report competitive or best-in-table accuracy on eight graph-classification benchmarks plus OGBG-MOLHIV, and strong scaffold-split ROC-AUC on seven MoleculeNet tasks, together with runtime comparisons against PH, PersLay, and TopoGCL and ablations versus PH-MLP/PH-TR, window width, and multi-filtration choices.

Significance. If the empirical gains are genuinely driven by the sequential interlevel representation, the work offers a practical, parallelizable bridge between TDA and attention models that avoids full persistence diagrams and vectorization design. Strengths that should be credited include: (i) a clean stability reduction to standard algebraic stability of interlevel modules with an explicit proof sketch in Appendix B; (ii) broad empirical coverage (classification, OGBG-MOLHIV, MoleculeNet) with multiple filtration families and runtime tables; (iii) public code; and (iv) a useful control (PH-TR) showing that treating Betti curves as sequences already helps. The contribution is incremental rather than foundational, but it is a credible and usable design pattern for topology-aware graph transformers.

major comments (4)
  1. [§4.3, Table 5, Table 15] §4.3 / Table 5 and the extended Table 15: the central attribution claim—that gains come from Topo-Scan’s sliding-window sequential tokens rather than from Transformer capacity or filtration choice—is only partially isolated. PH-TR (identical filtrations and Betti information, same Transformer backend) already closes most of the gap over PH-MLP; TopoFormer adds a further ~1–3 points on average. On several datasets the TopoFormer–PH-TR difference sits inside reported standard deviations (e.g., PROTEINS, IMDB-B under degree). A load-bearing control is missing: e.g., (a) shuffled or randomly ordered Topo-Scan tokens into the same Transformer, and/or (b) a fixed Transformer trained on non-topological fixed-length sequences of comparable dimension. Without such a control, the claim that the interlevel sequential structure itself is the driver remains under-supported.
  2. [§C.5, Table 6, Table 13] §C.5 / Table 6: window width m is selected as m=2 after experiments on the same graph-classification families later used for SOTA reporting. The multi-filtration menu (degree, Ollivier–Ricci, HKS) is likewise tuned on those families (§C.5, Table 13). This is ordinary hyperparameter practice, but it weakens the claim that the method is ‘lightweight’ and that gains transfer without dataset-specific topological engineering. Please either (i) fix m and the filtration menu a priori (or via a held-out meta-validation split) and re-report main tables, or (ii) clearly mark main results as post-selection and add a frozen-hyperparameter column.
  3. [Tables 2–4, Table 7, Abstract, §4.2] Tables 2–3 and Table 4: SOTA / ‘best or second-best on 7/8’ language should be tempered where margins are smaller than cross-fold standard deviations (MUTAG 94.68±4.30; BZR 92.36±4.11) and where TopoFormer* (not pure TopoFormer) is the reported model. On OGBG-MOLHIV, TopoFormer* reaches 78.19±0.19 versus Graphormer 80.51±0.53 and TopER 80.21±0.15; calling this ‘underscoring scalability and strength of topological signals’ overstates a ~2-point deficit under fingerprint fusion. For MPP (Table 3), pure topological performance is secondary to ECFP fusion (Table 7); the abstract and §4.2 should separate pure TopoFormer from TopoFormer* more sharply when claiming competitiveness with KANO/MV-Mol.
  4. [§3.1, Theorem 3.1, Corollary 3.2, Remark B.3] §3.1 / Theorem 3.1 and Corollary 3.2: the stability result is correct as a discrete sampling bound on interlevel Betti numbers, but the constant C = C(bG, {αi}, m, s) is left existential and can grow with the number of windows and bar complexity. For the ‘predictable and efficient compute’ and ‘robust encodings’ claims in the abstract, either give a concrete bound in terms of T and the number of bars, or clarify that stability alone does not imply that the four-dimensional token retains task-relevant information (Remark B.3 already notes the expressivity gap). A short remark on information loss relative to full barcodes would keep the theory claim proportionate.
minor comments (6)
  1. [Figure 3] Figure 3 caption and body: β0/β1 sequences are written with angle brackets in a way that can be confused with inner products; use parentheses or explicit tuples consistently.
  2. [§3] §3: notation switches between bGi, bG, and script G; unify clique-complex notation early.
  3. [Table 2] Table 2 header marks 1st/2nd/3rd by color; ensure a non-color cue (bold/underline) for accessibility and print.
  4. [§4.1] §4.1 Computational Complexity: the claim ‘no cycle-rank identity due to clique complexes’ is slightly opaque; one sentence explaining why β1 needs edge–triangle operators would help non-TDA readers.
  5. [§2, Remark B.3] Related work: Graph Filtration Learning (Hofer et al., 2020) is cited in the appendix remark but deserves a clearer main-text contrast (learn filtration vs. fix filtration and change representation).
  6. [Abstract, title block] Typos / style: ‘Topoformer’ vs ‘TOPOFORMER’ vs ‘Topo-Scan’ capitalization is inconsistent in the abstract and title block; ‘early-saturate’ hyphenation; arXiv id line appears inside the manuscript body.

Circularity Check

0 steps flagged

No significant circularity: empirical method with external benchmarks and an independent stability proof.

full rationale

TopoFormer is a methods/empirical paper. Topo-Scan is defined as sliding-window interlevel slices with tokens (β0, β1, |Vi|, |Ei|); that definition does not encode classification labels or force reported accuracies. Stability (Theorem 3.1, Corollary 3.2, Appendix B) reduces tokens’ ℓ1 change to bottleneck distance of interlevel modules and cites external algebraic-stability results (Botnan & Lesnick; Bauer & Lesnick; Cohen-Steiner et al.), not author-only uniqueness claims that forbid alternatives. Performance claims are measured on public benchmarks against published baselines (Tables 2–5, 8, 12–15), with ablations (PH-MLP vs PH-TR vs TopoFormer; width m; single vs multi-filtration) that do not rename a fitted parameter as a prediction. Remark B.3 explicitly frames Topo-Scan as a coarse sampling of the rank invariant rather than smuggling a known barcode result under a new name as a forced derivation. Same-group citations (e.g., TopER as a baseline; the Coskunuzer–Akçora tutorial for PH background) are ordinary and not load-bearing for the accuracy or stability claims. Hyperparameter choices (m=2 after Table 6) are standard model selection, not circular prediction-by-construction. No self-definitional loop, fitted-input-as-prediction, or uniqueness-import chain is present.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

Load-bearing content is mostly standard algebraic topology plus engineering choices. Stability rests on published interlevel-module bottleneck bounds and interval-rank Lipschitzness on a fixed clique complex. Empirical claims rest on hand-chosen filtration families, window schedule (m, N, stride), and architecture hyperparameters selected via experiments on the evaluation datasets. No new physical entity is postulated; Topo-Scan/TopoFormer are algorithmic constructs.

free parameters (4)
  • window width m = m=2 (default)
    Controls slice thickness; selected as m=2 after comparing m∈{2,3,4} on the classification datasets (Table 6, §C.5).
  • number of thresholds N / sequence length = N≈20 (main experiments)
    Sets token sequence length (20 thresholds → length-19 sequences, 76 features per filtration in main setup).
  • filtration function menu = task-specific fixed set
    Degree, Ollivier–Ricci, HKS for classification; atomic weight + Ricci for molecules—chosen as task-aligned scalars, not learned end-to-end in the main model.
  • transformer hidden size, dropout, LR, fusion weights α,β = hidden=32, dropout=0.5, lr=0.001 (classification)
    Standard training hyperparameters (hidden 32, dropout 0.5, Adam 1e-3, etc.) and learnable modality weights in the dual-transformer/MLP fusion.
axioms (4)
  • standard math Interlevel (level-set) persistence modules of upper-star filtrations on a fixed complex are stable in bottleneck distance with d_B(M^f, M^g) ≤ ||f−g||_∞ (Botnan & Lesnick).
    Lemma B.1 / Corollary 3.2; underpins the stability guarantee for Topo-Scan tokens.
  • standard math Interval-decomposable p.f.d. modules have interval-rank that changes only for bars near interval endpoints under a δ-matching (Bauer–Lesnick; Bjerkevik).
    Lemma B.2 used to bound per-window Betti differences and sum to discrete ℓ1 stability (Thm 3.1).
  • domain assumption Clique-complex 2-skeleton with upper-star extension of node functions is an adequate domain for graph topology in learning tasks.
    Throughout §3 and App. B; standard in graph PH but discards higher cliques beyond triangles for β1 computation practice.
  • ad hoc to paper Strict nested sublevel filtrations can be replaced by overlapping interlevel slices without losing the useful multi-scale signal for downstream Transformers.
    Core modeling claim in §3 motivating Topo-Scan vs PH early saturation; supported empirically (Figs. 2, 6–7) but not proved optimal.
invented entities (2)
  • Topo-Scan independent evidence
    purpose: Decompose a graph into a short ordered sequence of interlevel topological tokens (β0, β1, |V|, |E|) via sliding windows on a filtration.
    Central algorithmic object; defined operationally, not as a new mathematical space.
  • TopoFormer / TopoFormer* independent evidence
    purpose: Transformer (plus optional fingerprint MLP fusion) that consumes Topo-Scan sequences for graph-level prediction.
    Architecture wrapper; empirical artifact evaluated on public benchmarks.

pith-pipeline@v1.2.0-daily-grok45 · 34602 in / 3773 out tokens · 83224 ms · 2026-07-31T13:02:04.449835+00:00 · methodology

0 comments
read the original abstract

We introduce Topoformer, a lightweight and scalable framework for graph representation learning that encodes topological structure into attention-friendly sequences. At the core of our method is Topo-Scan, a novel module that decomposes a graph into a short, ordered sequence of topological tokens by slicing over node or edge filtrations. These sequences capture multi-scale structural patterns, from local motifs to global organization, and are processed by a Transformer to produce expressive graph-level embeddings. Unlike traditional persistent homology pipelines, Topo-Scan is parallelizable, avoids costly diagram computations, and integrates seamlessly with standard deep learning architectures. We provide theoretical guarantees on the stability of our topological encodings and demonstrate state-of-the-art performance across graph classification and molecular property prediction benchmarks. Our results show that Topoformer matches or exceeds strong GNN and topology-based baselines while offering predictable and efficient compute. This work opens a new path for parallelizable and unifying approaches to graph representation learning that integrate topological inductive biases into attention frameworks.

Figures

Figures reproduced from arXiv: 2607.28259 by Astrit Tola, Baris Coskunuzer, Cuneyt Gurcan Akcora, Md Joshem Uddin.

Figure 1
Figure 1. Figure 1: Topo-Scan. Topo-Scan decomposes a graph into sequential topological slices via node and edge filtrations. The top row shows node-based filtrations; the bottom row, edge-based ones. While persistence diagram computation is standardized, vectorization remains ad hoc and significantly impacts model perfor￾mance (Ali et al., 2022). Our goal is to develop an efficient and scalable alternative to this workflow f… view at source ↗
Figure 3
Figure 3. Figure 3: Topo-Scan vs. PH. This toy example highlights the key differences between the Topo-Scan filtration and standard PH filtration, where node values indicate filtration function values. In PH, early-activated nodes quickly saturate the graph, suppressing the emergence of later topological features. As shown, PH yields relatively uninformative barcodes with β0 = ⟨1, 2, 1, 1, 1⟩ and β1 = ⟨0, 0, 0, 0, 0⟩, while T… view at source ↗
Figure 2
Figure 2. Figure 2: PH vs. Topo-Scan. Average Betti-0 counts under degree filtration with 100 thresholds on IMDB-B and IMDB￾M. Standard PH saturates early, causing a sharp decline near the end of the curve. Topo-Scan avoids early saturation and continues to reveal late-emerging features, often surpassing PH counts at higher thresholds. The resolution (N) determines the number of slices, while the thickness (m) speci￾fies the … view at source ↗
Figure 4
Figure 4. Figure 4: TopoFormer Flowchart: Given an input graph G, sequential substructures are extracted via Topo-Scan. Each substructure is encoded into a four-dimensional topological signature. These sequences are processed by a transformer model, and outputs from multiple filtration functions are fused using attention-based concatenation. A final prediction layer maps the representation to the target graph property. Coroll… view at source ↗
Figure 5
Figure 5. Figure 5: TOPOFORMER∗ : To successfully integrate complementary graph information, such as ECFP, with our TOPOFORMER model, we employ a MLP. The MLP output is combined with the TOPOFORMER model using an attention mechanism, and the combined representation is then passed through a graph prediction network to perform the final prediction task [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: PH vs. Topo-Scan. Normalized average Betti-0 values over 20 thresholds of the degree-centrality filtration on (a) BZR, (b) MUTAG, and (c) PROTEINS. Under the classical PH pipeline, feature counts decline monotonically with increasing threshold, whereas Topo-Scan maintains elevated values at higher thresholds, revealing late-emerging topological features that PH alone misses. threshold range, new features c… view at source ↗
Figure 7
Figure 7. Figure 7: PH vs. Topo-Scan. Bar plots of Betti-1 counts at each O.Ricci filtration threshold on the PROTEINS dataset, with bar color intensity encoding the frequency of each integer value at that threshold. In (a) classical PH features rapidly taper off and plateau early, whereas in (b) Topo-Scan shows increasingly darker bars at higher thresholds, evidence of continued cycle emergence beyond PH’s saturation. 24 [P… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 8 linked inside Pith

  1. [1]

    Two stability lemmas

    We write dB(M f k , Mg k ) for the bottleneck distance between the interval decompositions (barcodes) of the interlevel modulesM f k andM g k . Two stability lemmas. Lemma B.1(Interlevel stability).(Botnan & Lesnick, 2018, Thm 1.1 & 1.2) For k≥0 , the interlevel modules of the upper–star filtrations induced by f, g:V→R on the fixed clique complex bG satis...

    2018

  2. [2]

    The outputs z1, z2, and z3 are then combined through a learnable weighted sum

    Finally, the output of the MLPMis denotedz 3 ∈R B×H . The outputs z1, z2, and z3 are then combined through a learnable weighted sum. Specifically, the combined feature vectorz combined is computed as: zcombined =α·z 1 +β·z 2 + (1−α−β)·z 3 where α and β are learnable parameters that control the contribution of each modality to the final representation. Thi...

    2026

  3. [5]

    For the MLP component, we employed a two-layer MLP with a hidden dimension of 200, ensuring that its output dimension matches the output dimension of the TOPOFORMERmodel

    In our hybrid model, we used the same experimental setup for the TOPOFORMERcomponent. For the MLP component, we employed a two-layer MLP with a hidden dimension of 200, ensuring that its output dimension matches the output dimension of the TOPOFORMERmodel. The model was optimized using the Adam optimizer with a learning rate of 0.01 and a weight decay of ...

    2021

  4. [6]

    and (ii)unnormalizedBetti-0 when PH and Topo-Scan share similar ranges (Fig. 2). Betti-1 frequency barplots are shown to illustrate higher-order behavior (Fig. 7). How to read the figures.A positive Betti-0 value at a position means additional connected components are present in that slice/threshold; persistent nonzero values toward theright sideof the ho...

    2026

  5. [10]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929,

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, and Sylvain Gelly. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929,

    Pith/arXiv arXiv 2010

  6. [11]

    Knowledge graph-enhanced molecular contrastive learning with functional prompt

    Yin Fang, Qiang Zhang, Ningyu Zhang, Zhuo Chen, Xiang Zhuang, Xin Shao, Xiaohui Fan, and Huajun Chen. Knowledge graph-enhanced molecular contrastive learning with functional prompt. Nature Machine Intelligence, 5(5):542–553, 2023a. Zhongxi Fang, Jianming Huang, Xun Su, and Hiroyuki Kasai. Wasserstein graph distance based on l1–approximated tree edit dista...

    2083

  7. [12]

    Gemnet: Universal directional graph neural networks for molecules.Advances in neural information processing systems, 34:6790–6802,

    12 Published as a conference paper at ICLR 2026 Johannes Gasteiger, Florian Becker, and Stephan Günnemann. Gemnet: Universal directional graph neural networks for molecules.Advances in neural information processing systems, 34:6790–6802,

    2026

  8. [13]

    Smiles transformer: Pre-trained molecular fingerprint for low data drug discovery.arXiv preprint arXiv:1911.04738,

    Shion Honda, Shoi Shi, and Hiroki R Ueda. Smiles transformer: Pre-trained molecular fingerprint for low data drug discovery.arXiv preprint arXiv:1911.04738,

    Pith/arXiv arXiv 1911

  9. [14]

    N-gram graph: Simple unsupervised repre- sentation for graphs, with applications to molecules.Advances in neural information processing systems, 32,

    13 Published as a conference paper at ICLR 2026 Shengchao Liu, Mehmet F Demirel, and Yingyu Liang. N-gram graph: Simple unsupervised repre- sentation for graphs, with applications to molecules.Advances in neural information processing systems, 32,

    2026

  10. [15]

    Pre- training molecular graph representation with 3d geometry

    Shengchao Liu, Hanchen Wang, Weiyang Liu, Joan Lasenby, Hongyu Guo, and Jian Tang. Pre- training molecular graph representation with 3d geometry. InICLR Workshop on Geometrical and Topological Representation Learning, 2022a. Yi Liu, Limei Wang, Meng Liu, Yuchao Lin, Xuan Zhang, Bora Oztekin, and Shuiwang Ji. Spherical message passing for 3d molecular grap...

    2082

  11. [16]

    Graph alignment kernels using weisfeiler and leman hierarchies

    Giannis Nikolentzos et al. Graph alignment kernels using weisfeiler and leman hierarchies. In International Conference on Artificial Intelligence and Statistics, pp. 2019–2034. PMLR,

    2019

  12. [20]

    Model-agnostic augmentation for accurate graph classification

    Jaemin Yoo et al. Model-agnostic augmentation for accurate graph classification. InProceedings of the ACM Web Conference 2022, pp. 1281–1291,

    2022

  13. [21]

    Benchmarking large language models for molecule prediction tasks.arXiv:2403.05075,

    Zhiqiang Zhong, Kuangyu Zhou, and Davide Mottin. Benchmarking large language models for molecule prediction tasks.arXiv:2403.05075,

    Pith/arXiv arXiv

  14. [25]

    Proof of Theorem 3.1.By Theorem 3.1, we have∥ bβk(G, f)− bβk(G, g)∥1 ≤C d B(M f k , Mg k )

    Theorem 3.2.For upper–star filtrations on a fixed complex, dB(M f k , Mg k )≤ ∥f−g∥∞ (Lemma B.1), hence∥ bβk(G, f)− bβk(G, g)∥1 ≤C∥f−g∥ ∞. Proof of Theorem 3.1.By Theorem 3.1, we have∥ bβk(G, f)− bβk(G, g)∥1 ≤C d B(M f k , Mg k ). By Lemma B.1 (interlevel/level-set stability on the fixed clique complex),dB(M f k , Mg k )≤ ∥f−g∥∞. Combining the two inequal...

    2007

  15. [26]

    A full expressivity comparison and formal information-loss bounds relative to complete barcodes are interesting directions for future work

    can be seen as learning the filtration function h, while our work fixes h and instead changes the representation from global barcodes to local interlevel sequences. A full expressivity comparison and formal information-loss bounds relative to complete barcodes are interesting directions for future work. C MORE ONTOPOFORMER C.1 BASEMODEL: TRANSFORMER Our T...

    2026

  16. [28]

    applies a Haar wavelet transform to graph signals and performs pooling by selecting key wavelet coefficients. Our model TOPOFORMERtakes a different approach by integrating multiscale topological filtrations with a transformer-based attention mechanism, enabling the pooling of substructures across scales and yielding robust higher-order graph representatio...

    2026

  17. [29]

    from three to seven filtration functions and compares three topological pipelines under the same filtration function: the classical PH-MLP baseline (sublevel PH + Betti vector + MLP), our PH-TR variant (same Betti vectors but processed as sequences by a Transformer), and TOPOFORMER(Topo-Scan sequences with sliding-window interlevel filtrations). Across al...

    2026

  18. [2005]

    They encode features such as atomic connectivity and substructural patterns, enabling efficient similarity search and predictive modeling

    16 Published as a conference paper at ICLR 2026 Appendix A TOPOFORMER ∗: TOPOFORMER FORMPP A.1 MOLECULARFINGERPRINTS Molecular fingerprints are widely used in computational chemistry and machine learning to represent molecular structures as fixed-length numerical vectors (Cereto-Massagué et al., 2015). They encode features such as atomic connectivity and ...

    2026

  19. [2007]

    Gabriele Corso, Luca Cavalleri, Dominique Beaini, Pietro Liò, and Petar Veli ˇckovi´c

    doi: 10.1007/s00454-006-1276-5. Gabriele Corso, Luca Cavalleri, Dominique Beaini, Pietro Liò, and Petar Veli ˇckovi´c. Principal neighbourhood aggregation for graph nets.Advances in neural information processing systems, 33: 13260–13271,

    doi:10.1007/s00454-006-1276-5

  20. [2010]

    Self-supervised graph transformer on large-scale molecular data.Advances in neural information processing systems, 33:12559–12571,

    14 Published as a conference paper at ICLR 2026 Yu Rong, Yatao Bian, Tingyang Xu, Weiyang Xie, Ying Wei, Wenbing Huang, and Junzhou Huang. Self-supervised graph transformer on large-scale molecular data.Advances in neural information processing systems, 33:12559–12571,

    2026

  21. [2014]

    Topogcl: Topological graph contrastive learning

    Yuzhou Chen, Jose Frias, and Yulia R Gel. Topogcl: Topological graph contrastive learning. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pp. 11453–11461, 2024a. 11 Published as a conference paper at ICLR 2026 Yuzhou Chen, Ignacio Segovia-Dominguez, Cuneyt Gurcan Akcora, Zhiwei Zhen, Murat Kantar- cioglu, Yulia Gel, and Baris...

    2026

  22. [2017]

    Persistence homology of networks: methods and applications.Applied Network Science, 4(1):1–28,

    10 Published as a conference paper at ICLR 2026 Mehmet E Aktas, Esra Akbas, and Ahmed El Fatmaoui. Persistence homology of networks: methods and applications.Applied Network Science, 4(1):1–28,

    2026

  23. [2018]

    Sparks of artificial general intelligence: Early experiments with gpt-4.arXiv preprint arXiv:2303.12712,

    Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt-4.arXiv preprint arXiv:2303.12712,

    Pith/arXiv arXiv

  24. [2019]

    A survey of vectorization methods in topological data analysis.arXiv preprint arXiv:2212.09703,

    Dashti Ali, Aras Asaad, Maria-Jose Jimenez, Vidit Nanda, Eduardo Paluzo-Hidalgo, and Manuel Soriano-Trigueros. A survey of vectorization methods in topological data analysis.arXiv preprint arXiv:2212.09703,

    Pith/arXiv arXiv

  25. [2020]

    Topological methods in machine learning: A tutorial for practitioners.arXiv preprint arXiv:2409.02901,

    Baris Coskunuzer and Cüneyt Gürcan Akçora. Topological methods in machine learning: A tutorial for practitioners.arXiv preprint arXiv:2409.02901,

    Pith/arXiv arXiv

  26. [2021]

    Moleculenet: a benchmark for molecular machine learning

    15 Published as a conference paper at ICLR 2026 Zhenqin Wu, Bharath Ramsundar, Evan N Feinberg, Joseph Gomes, Caleb Geniesse, Aneesh S Pappu, Karl Leswing, and Vijay Pande. Moleculenet: a benchmark for molecular machine learning. Chemical science, 9(2):513–530,

    2026

  27. [2022]

    Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805,

    Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805,

    Pith/arXiv arXiv

  28. [2023]

    Understanding the power of persistence pairing via permutation test.arXiv preprint arXiv:2001.06058,

    Chen Cai and Yusu Wang. Understanding the power of persistence pairing via permutation test.arXiv preprint arXiv:2001.06058,

    Pith/arXiv arXiv 2001

  29. [2024]

    Ripsnet: a general architecture for fast and robust estimation of the persistent homology of point clouds

    Thibault de Surrel, Felix Hensel, Mathieu Carrière, Théo Lacombe, Yuichi Ike, Hiroaki Kurihara, Marc Glisse, and Frédéric Chazal. Ripsnet: a general architecture for fast and robust estimation of the persistent homology of point clouds. InTopological, algebraic and geometric learning workshops 2022, pp. 96–106. PMLR,

    2022

  30. [2025]

    Communicative representation learning on attributed molecular graphs

    Ying Song, Shuangjia Zheng, Zhangming Niu, Zhang-Hua Fu, Yutong Lu, and Yuedong Yang. Communicative representation learning on attributed molecular graphs. InIJCAI, volume 2020, pp. 2831–2838,

    2020