acceptodds
Under review as a conference paper at ICLR 2027

Learning From Filtration Sketches

Abstract

Methods from computational topology are widely used in machine learning to extract topological and geometric information from data. Their use, however, often involves several computational stages before a representation suitable for learning is obtained. We investigate whether the underlying mathematical object, a filtration of a simplicial complex built on the data, can itself serve as the basis of a learning representation. We introduce *filtration sketches*, which combine estimated spectral summaries at selected filtration levels with approximate persistent ranks between them. Both components are computed through sparse boundary-operator applications and randomized trace estimation, without explicit eigendecomposition or persistence-barcode computation. The resulting fixed-width sequence is processed by a sequence model for downstream learning. On molecular property prediction and point-cloud benchmarks, filtration sketches outperform topology-based baselines and match or outperform point-cloud-specific neural networks.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Persistent homology, Filtrations, Simplicial complexes

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 4 reviews, above 27% of the others.
Reviewer_CY6oOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies an interseting question: can the filtration itself provides more information rather than persistent homology, which requires more time to compute. The authors propose to build a sequence from the filtration and then naturally use a sequence model to downstream machine learing tasks. The sequence contains two major parts: hodgre stream built from the hodge laplacian of each simplical complex in the filtration and persistent rank stream for encoding the relation between different complexes. A third filtration coordinates are also considered. After obtaining the sequence, TCN is used for representation leanring. The proposed FiltSketch is tested on QM7b dataset for molecular property prediction and more traditional datasets like ModelNet10 for classification and regression tasks. Filtsketch is shown to have similar performance with non-topological models on some datasets.

Critical strengths
  1. This paper proposes a principled way to learn directly from the filtrtation instead of the filtration->PH->vectorization pipeline. Filtsketch combines local Hodge stream with persistent rank as the sequence and use strandard sequence modeling technique. The hodge stream encodes spectral information and the persisent rank stream encodes the relation between complexes. The later one is an improvement from existing TopoFormer, which did not consider relational information explictly. This is a substantive methodological strength.

  2. The results seems promising. The experiments covers molecular property prediction and point cloud classification/regression. These results support the practical usefulness of the proposed combionation.

Critical weaknesses
  1. The computational advantage is not analyzed or empircally established. Avoiding exact persistent homology computation is an important motivation. But the paper did not report time complexity of report actual runtime, compared with exact persistent homology computation+vectorization. The peak memory or scaleup test are also not provided. The proposed Filtsketch still needs buliding simplicial complexes and appplying boundary operators. This leaves a central practical advangte insufficiently supported.

  2. The comparisons did not isolate the contribution of the sketch representation. Although the Flood PH baseline share the same filtration, they use differernt dienstream model for different vectorization methods: PI uses CNN, ATOL uses XGBoost and FiltSkectch uses TCN. And the FiltSketch on ModelNet10 dataset uses cubical filtration without the corresponding PH baselines. These experiments provided limited evidence for the advantage of Filtsketch.

  3. Insufficient comparison with TopoFormer. TopoForm is the cloest work, which uses Betti numbers on graph to build sequence and transformer for sequence modeling. It is a direct comparison of Filtsketch, which uses hodge and persistence rank stream to build sequence and TCN for sequence modeling. With fixed sequnce modeling method TCN/transformrt, the betti stream should be compared with the proposed stream. And with the same stream, different sequence modeling method should be considered.

two less minor points:

  1. On the QM7b benchmark, only topological methods (TopoFormer) and Davies (2023) are considered. The comparision with point cloud baselines are missing. And the point cloud baselines are rather old.

  2. The ablation study is rather limited on two datasets for two tasks. More datasets could make the reaults more reliable.

Questions for authors
  1. Can filtration learning method be combined into the proposed Filtsketch pipeline?
  2. On the only Rocks dataset dataset where topology method outperforms the non-topological point cloud baselines, can the author explain the reason behind this?
  3. Any stability analysis on the representation?
3 comments
AuthorsOpenReview ↗

Response to Reviewer CY6o (1/3)

W1: The computational advantage is not analyzed or empircally established. Avoiding exact persistent homology computation is an important motivation. But the paper did not report time complexity of report actual runtime, compared with exact persistent homology computation+vectorization. The peak memory or scaleup test are also not provided. The proposed Filtsketch still needs buliding simplicial complexes and applying boundary operators. This leaves a central practical advangte insufficiently supported.

We have now included a detailed runtime comparison between FiltSketch and PH+PI, as well as a breakdown of runtime across the different computational steps in Appendix E. For details, we kindly refer to our general comment to all reviewers on runtime and efficiency. Overall, our runtime study shows favorable scaling behavior of FiltSketch with smaller runtime than PH+PI when using the Flood complex and increasing the number of landmarks.


W2: The comparisons did not isolate the contribution of the sketch representation. Although the Flood PH baseline share the same filtration, they use differernt dienstream model for different vectorization methods: PI uses CNN, ATOL uses XGBoost and FiltSkectch uses TCN. And the FiltSketch on ModelNet10 dataset uses cubical filtration without the corresponding PH baselines. These experiments provided limited evidence for the advantage of Filtsketch.

We used the PI+CNN and ATOL+XGBoost combinations, as they typically show strong performance, but we understand that a more systematic analysis on the downstream architecture provides stronger evidence whether the benefits are due to the representation itself. We therefore compare classification performance on the MCB-C dataset for different classification models between the FiltSketch and PI representations. Specifically, we evaluate both on the same CNN architecture that was used for PI in the main experiments (with 2D convolutions replaced by 1D convolutions for FiltSketch) and on MLPs. As non neural network methods we use kk-NN, random forest, LightGBM and XGBoost. As additional sequence models for FiltSketch, we use a GRU and an LSTM. All models are trained on the same representations as in the original MCB-C experiment. Model specific hyperparameters are selected based on the validation accuracy on the first split (due to time constraints, we will tune on all splits and update the table).

ModelPIFiltSketch
kNN55.8 ± 3.958.1 ± 4.1
Random forest64.8 ± 2.065.0 ± 4.1
LightGBM69.5 ± 3.867.6 ± 4.1
XGBoost67.7 ± 2.567.7 ± 4.6
MLP60.9 ± 5.770.5 ± 3.5
CNN (2D/1D)66.2 ± 2.769.0 ± 3.5
GRU--69.7 ± 3.5
LSTM--63.4 ± 4.2
TCN--77.3 ± 3.0

We see two effects:

  • First, the standard machine learning methods achieve around the same balanced accuracy on both representations, but the neural networks (MLP and CNN) are able to leverage more information from FiltSketch than from the PIs, as they achieve better classification performances (69.0% vs 66.2% for CNN and 60.9% vs 70.5% for MLP), which suggests that performance benefits to some part can be attributed to the FiltSketch representations.
  • Second, parts of the performance gains of FiltSketch over PI are due to the classification model itself, or rather to how the classification model leverages the sequential information within the filtration sketches. The GRU model achieves similar performance to the MLP (69.7% vs. 70.5%), and the LSTM performs clearly worse, while the TCN achieves by far the best performance across all models (77.3%). To further assess this part, we performed an ablation over the different TCN building blocks, which we have also included in Appendix F and in our response to reviewer av2404.

We have also added a PH+PI baseline using the same cubical complex filtration as for FiltSketch on ModelNet10, where, for each filtration direction, we compute H0, H1 and H2 PI's and feed them through a ResNet-style CNN. As with FiltSketch, aggregation is done via a DeepSet module and we use the same MLP classification head. We have updated Table 2 with the results, which show that the PH+PI approach also benefits from the cubical filtration, yet classification balanced accuracy remains below that of our FiltSketch approach (84% vs. 92%).

Finally, we want to clarify that we included the cubical complex filtration not to demonstrate an improvement of FiltSketch over other PH methods (which benefit as well from this construction, as can be seen from the added PH + PI result), but to show (1) that FiltSketch works with filtrations other than the Flood complex filtration and (2) that one can indeed almost match the performance of neural baselines with a topological approach.

AuthorsOpenReview ↗

Response to Reviewer CY6o (2/3)

W3: Insufficient comparison with TopoFormer. TopoForm is the cloest work, which uses Betti numbers on graph to build sequence and transformer for sequence modeling. It is a direct comparison of Filtsketch, which uses hodge and persistence rank stream to build sequence and TCN for sequence modeling. With fixed sequnce modeling method TCN/transformrt, the betti stream should be compared with the proposed stream. And with the same stream, different sequence modeling method should be considered.

As requested, we performed a control experiment, where we extend the TopoFormer (Uddin et al., 2026) features (originally Betti-0/1 and vertex, edge counts on graphs) to contain Betti-0/1/2 (obtained from computing PH), as well as the number of vertices, edges, triangles and tetrahedra. We then supply this "feature stream" to our downstream TCN model. Results on Rocks and MCB are given below:

ModelFiltSketchTopoFormer-style features
Rocks97.2 ± 1.095.9 ± 1.0
MCB-C77.3 ± 3.075.4 ± 4.2

This shows that this type of feature stream is also competitive under exactly the same downstream (TCN) model but does not match the performance of FiltSketch.

For a comparison with fixed stream (FiltSketch) and different sequence modelling, we refer to our response to W1, where we compare with LSTM, GRU and the same CNN model that was used for the PI baseline but with 2D convolutions replaced by 1D convolutions. Moreover, below, we present an ablation over the different TCN components, to isolate which parts of our sequence modelling approach are relevant.

FiltSketchBal. acc. ↑#Param.
+ full (7-layer) TCN, original readout0.97 ± 0.011,588,235
+ 1-layer TCN (ks 3), original readout0.93 ± 0.02395,144
+ 1-layer TCN (ks 3), mean + linear readout0.82 ± 0.06261,639
+ 1-layer TCN (ks 1), mean + linear readout0.73 ± 0.07261,383

Overall, we can draw several conclusions from these results: First, the 1-layer TCN retains strong performance (0.93 versus 0.97 balanced accuracy) with approximately 75% fewer parameters, showin that the sketches remain predictive with a substantially shallower backbone. Second, replacing the original readout and nonlinear prediction head with mean pooling and a linear classifier reduces accuracy to 0.82, indicating that aggregation and prediction contribute considerably. Third, restricting the remaining convolution to kernel size one further reduces accuracy to 0.73 while removing only 256 parameters, supporting the value of local cross-timepoint interactions rather than parameter count alone.
In summary, the results demonstrate that useful predictive information is in fact accessible to simpler models, while also showing that the full downstream architecture improves its extraction.


W4 (minor): On the QM7b benchmark, only topological methods (TopoFormer) and Davies (2023) are considered. The comparision with point cloud baselines are missing. And the point cloud baselines are rather old.

In Davies et al., the authors additionally list two neural baselines, DTNN (Schütt et al., 2017) and GGRNN (Shindo & Matsumoto, 2019) which both perform better than the persistent Laplacians and also better than FiltSketch on QM7b. We deliberately only compare against approaches that leverage topological information, but will acknowledge that neural models tend to perform better on this dataset. We however hypothesize that once difference filtrations are combined on QM7b, one may be able to match the neural baselines (like on ModelNet10).


W5 (minor): The ablation study is rather limited on two datasets for two tasks. More datasets could make the reaults more reliable.

We deliberately chose two datasets for our ablation that are quite different in nature, i.e., a point cloud dataset where we use Flood complexes and a graph dataset where we use the supplied Coulomb-matrix filtration. The same two datasets are now also used to assess the impact of using exact persistence ranks vs. our spectral summaries in the persistence rank stream (see Appendix F, Table 5).

AuthorsOpenReview ↗

Response to Reviewer CY6o (3/3)

Q1: Can filtration learning method be combined into the proposed Filtsketch pipeline?

While we won't rule it out, combining filtration learning with FiltSketch is challenging, and we think that a modified approach where filtration learning is targeted from the start would be more promising than trying to adjust the FiltSketch pipeline. For the latter, we see two immediate obstacles.

  • First, we currently preselect the filtration levels tt, and the sketch depends on the filtration function ff only through the indicator functions 1f(σ)≤t1_{f(\sigma) \le t}, which are piecewise constant so gradients are zero. In principle, it would be possible to replace them by soft indicators, similar to what is done in (Differentiable Euler Characteristic Transforms for Shape Classification, Röell and Rieck, ICLR 2024), but one would need to look into that.
  • Second, computational cost. We would need to backpropagate through the FiltSketch computation, whereas for PH the gradients are given by the persistence pairs, so no backpropagating through the PH computation is necessary.

Q2: On the only Rocks dataset dataset where topology method outperforms the non-topological point cloud baselines, can the author explain the reason behind this?

The neural network baselines fail on the rocks classification task (~50% accuracy on the binary problem). The underlying reason is the nature of the data. The dataset consists of large, dense point clouds containing up to 16 million points that simulate porous media, and was introduced by Graf et al. (NeurIPS 2026) to demonstrate the failure of subsampling based approaches; The class-specific information lies in the pores, i.e., unoccupied regions in the point cloud, but subsampling can only recognize unfilled regions that are larger than the sampling resolution.

The topology methods use Flood complexes to capture the full point cloud, but the neural networks must operate on a substantially smaller subsets due to memory and runtime constraints (following Graf et al., we uniformly sample 2k points).


Q3: Any stability analysis on the representation?

The Hodge stream itself is not stable, i.e., for filtration functions ff and gg, there is no bound of the form ∥h(f)−h(g)∥≤C∥f−g∥∞\|h(f)-h(g)\| \le C \|f-g\|_\infty, where hh denotes the concatenated Hodge features. The reason is the same as for homology at isolated filtration levels, i.e., small perturbations in the filtration values of the simplices could move many simplices across a sampled filtration level, and thereby produce large changes in the boundary matrices at that level.
The same reasoning applies to the rank stream, as it also relies on comparing two specific filtration levels, whose boundary matrix can change to a large extent. Persistent homology instead records the filtration values where Betti numbers (and persistent ranks) change, and the multiset of all birth and death pairs depends Lipschitz-continuously on the filtration function.

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

The paper proposes FiltSketch, a representation for learning from filtrations without explicitly computing persistence barcodes or eigendecompositions. It combines local Hodge spectral summaries with approximate persistent ranks between filtration levels, estimated through sparse matrix operations and randomized trace estimation. A temporal convolutional network processes these features for molecular property prediction and point-cloud tasks, achieving promising improvements over several topology-based baselines.

Critical strengths
  1. The motivation is meaningful. Exact persistence barcodes need not be the best intermediate representation for every downstream task. Combining topology with additional spectral information is a reasonable direction.
  2. The empirical results are encouraging.
Critical weaknesses
  1. Computational efficiency is not demonstrated experimentally. Although efficiency motivates the method, the paper reports no runtime, memory, or scaling comparisons against PH and persistent-Laplacian baselines. The complexity analysis alone does not establish a practical advantage.
  2. The experiments do not isolate the representation’s contribution. Different representations use different downstream predictors. Comparisons against exact persistent ranks and Betti-number/simplex-count sequences using the same TCN would help determine whether the proposed descriptors explain the improvements.
  3. The improved ModelNet10 result lacks matched-filtration baselines. Switching to directional cubical filtrations raises accuracy from 0.74 to 0.92, but PH baselines are evaluated only with Flood filtrations. The experiment therefore does not separate the benefit of filtration choice from that of the proposed representation.
Questions for authors
  1. Could the authors provide runtime, peak-memory, and scaling comparisons against PH-based and persistent-Laplacian baselines?
  2. How much of the performance improvement comes from the proposed filtration descriptors versus the downstream TCN?
  3. On ModelNet10, could the authors evaluate PH-based baselines using the same directional cubical filtrations and directional views as FiltSketch?
1 comment
AuthorsOpenReview ↗

Response to Reviewer LWYj

W1: Computational efficiency is not demonstrated experimentally. Although efficiency motivates the method, the paper reports no runtime, memory, or scaling comparisons against PH and persistent-Laplacian baselines. The complexity analysis alone does not establish a practical advantage.

We agree that a complexity analysis alone does not establish a practical advantage, and refer the reviewer to our general comment section "Runtime and efficiency". In summary, we have now included a detailed runtime comparison (Appendix E) between FiltSketch and PH+PI, as well as a runtime breakdown of the different computational steps. Overall, our runtime study shows favorable scaling behavior of FiltSketch with smaller runtime than PH+PI when using the Flood complex and increasing the number of landmarks.


W2: The experiments do not isolate the representation’s contribution. Different representations use different downstream predictors. Comparisons against exact persistent ranks and Betti-number/simplex-count sequences using the same TCN would help determine whether the proposed descriptors explain the improvements.

As requested by multiple reviewers, we have added an experiment (Table 5, Appendix) where we replace the rank stream with exact persistence ranks. On the Rocks dataset balanced accuracy drops from 0.96 ± 0.01 to 0.93 ± 0.02, while on QM7b, both variants yield similar performance (MAE ratio of 0.72 vs. 0.73; lower is better). This indicates that our spectral summaries capture information beyond exact persistence ranks and that this information is beneficial downstream. We could also extend this analysis to the Hodge stream, should the reviewer consider it important.

In addition to that, we ablate the effect of different downstream architectures on top of the FiltSketch and PI representations for the MCB-C classification balanced accuracy in the following table (now Table 6, see Appendix F for additional implementation details).

ModelPIFiltSketch
kNN55.8 ± 3.958.1 ± 4.1
Random forest64.8 ± 2.065.0 ± 4.1
LightGBM69.5 ± 3.867.6 ± 4.1
XGBoost67.7 ± 2.567.7 ± 4.6
MLP60.9 ± 5.770.5 ± 3.5
CNN(2D/1D)66.2 ± 2.769.0 ± 3.5
GRU--69.7 ± 3.5
LSTM--63.4 ± 4.2
TCN--77.3 ± 3.0

We see two effects:

  • First, the standard machine learning methods achieve around the same balanced accuracy on both representations, but the neural networks (MLP and CNN) are able to leverage more information from FiltSketch than from the PIs, as they achieve better classification performances (69.0% vs 66.2% for CNN and 60.9% vs 70.5% for MLP), which suggests that performance benefits to some part can be attributed to the FiltSketch representations.
  • Second, parts of the performance gains of FiltSketch over PI are due to the classification model itself, or rather to how the classification model leverages the sequential information within the filtration sketches. The GRU model achieves similar performance to the MLP (69.7% vs. 70.5%), and the LSTM performs clearly worse, while the TCN achieves by far the best performance across all models (77.3%).

W3: The improved ModelNet10 result lacks matched-filtration baselines. Switching to directional cubical filtrations raises accuracy from 0.74 to 0.92, but PH baselines are evaluated only with Flood filtrations. The experiment therefore does not separate the benefit of filtration choice from that of the proposed representation.

We have now added a matched filtration baseline using the same cubical complex filtration as for FiltSketch where, for each filtration direction, we compute H0, H1 and H2 PI's and feed them through a ResNet-style CNN. As with FiltSketch, aggregation is done via a DeepSet module and we use the same MLP classification head. We updated Table 2 with the corresponding results, which show that the PH+PI approach also benefits from the cubical filtration, yet classification balanced accuracy remains below that of our FiltSketch approach (84% vs. 92%).

We also want to clarify that we included the cubical complex filtration not to demonstrate an improvement of FiltSketch over other PH methods (which benefit as well from this construction, as can be seen from the added PH + PI result), but to show (1) that FiltSketch works with filtrations other than the Flood complex filtration and (2) that one can indeed almost match the performance of neural baselines with a topological approach.

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

The paper introduces FiltSketch, a representation combining local Hodge diffusion moments with approximate persistent ranks between sampled filtration levels. Both components use sparse boundary-operator products and randomized trace estimation, avoiding explicit eigendecomposition and barcode computation. A gated temporal convolutional network processes the resulting sequences. Experiments cover molecular property prediction and point-cloud tasks and improve over several topology-based baselines.

Critical strengths
  1. The construction provides a coherent connection between boundary operators, spectral summaries, and persistent ranks. The distinction between finite-depth approximation, stochastic trace estimation, and subsequent feature processing is explained carefully.
  2. The empirical evaluation spans multiple tasks and filtration constructions, uses repeated splits, and tunes the PH vectorization baselines. The QM7b results and strong Rocks performance provide encouraging evidence that the proposed descriptors are useful for learning.
Critical weaknesses
  1. Insufficient validation of the persistent-rank approximation. The experiments use only shallow diffusion depths and eight probes, without evaluating rank approximation error, clipping frequency, or sensitivity to these choices. Consequently, the rank-only ablation does not establish whether the gains reflect topological survival or other spectral and size-dependent signals. Comparisons with exact ranks and simple count features under the same TCN are needed.
  2. Key comparisons do not isolate the representation’s contribution. The QM7b experiment uses inverse-Coulomb filtration values, whereas the cited baseline describes Coulomb values, creating an apparent mismatch that requires clarification. Meanwhile, the ModelNet10 improvement also changes the filtration and introduces DeepSet aggregation, without corresponding PH controls. These comparisons leave the gains attributable specifically to FiltSketch unclear.
  3. The computational advantage remains unverified. Despite efficiency being a central motivation, the paper reports no runtime, peak-memory, or empirical scaling comparisons. The complexity bound excludes filtration construction and involves thousands of filtration-level pairs. Measurements against optimized PH and spectral baselines are necessary to establish the practical accuracy–cost tradeoff.
Questions for authors
  1. How do rank approximation error and clipping frequency vary with diffusion depth and probe count? How does an exact-rank stream perform with the same TCN?
  2. Were the Davies et al. baselines recomputed using inverse-Coulomb filtrations? Are results available for PH baselines using the same directional cubical inputs?
  3. What are the feature-extraction time, total preprocessing time, and peak memory relative to the principal baselines?
2 comments
AuthorsOpenReview ↗

Response to Reviewer ttLc (1/2)

W1: Insufficient validation of the persistent-rank approximation. The experiments use only shallow diffusion depths and eight probes, without evaluating rank approximation error, clipping frequency, or sensitivity to these choices. Consequently, the rank-only ablation does not establish whether the gains reflect topological survival or other spectral and size-dependent signals. Comparisons with exact ranks and simple count features under the same TCN are needed.

To address your concern, we have now added experiments (using the same TCN model) on Rocks and QM7b using exact persistence ranks (Table 5, Appendix F). On the Rocks dataset balanced accuracy drops from 0.96 ± 0.01 to 0.93 ± 0.02, while on QM7b, both variants yield similar performance (MAE ratio of 0.72 vs. 0.73; lower is better).

Overall, we think our current results support the usefulness of the spectral summaries. We, however, explicitly acknowledge that the variants differ in channel count (for the downstream TCN model) and coefficient field, and that size/degree effects may also contribute.

We have additionally added a general comment "On persistent rank estimates vs. exact persistence ranks" in our response to all reviewers which clarifies that we are more interested in spectral summaries rather than in accurate persistence rank estimates and that we will adjust the wording accordingly.


W2a: Key comparisons do not isolate the representation’s contribution. The QM7b experiment uses inverse-Coulomb filtration values, whereas the cited baseline describes Coulomb values, creating an apparent mismatch that requires clarification.

You are correct and we apologize for the inconsistency. We reproduced the results of the Davies et al. almost exactly using the provided Coulomb values, but our FiltSketch approach indeed used inverse-Coulomb filtration values. We re-ran those experiments, with the result that FiltSketch got even slightly better. Results are now in the updated Table 1 in the PDF. Thank you very much for catching this.

W2b: Meanwhile, the ModelNet10 improvement also changes the filtration and introduces DeepSet aggregation, without corresponding PH controls. These comparisons leave the gains attributable specifically to FiltSketch unclear.

As requested, we have now added a PH+PI baseline using the same cubical complex filtration as for FiltSketch where, for each filtration direction, we compute H0, H1 and H2 PI's and feed them through a ResNet-style CNN. As with FiltSketch, aggregation is done via a DeepSet module and we use the same MLP classification head.
We have added the corresponding result to the updated Table 2, which shows that the PH+PI approach also benefits from the cubical filtration, yet classification balanced accuracy remains below that of our FiltSketch approach (84% vs. 92%).

We also want to clarify that we included the cubical complex filtration not to demonstrate an improvement of FiltSketch over other PH methods (which benefit as well from this construction, as can be seen from the added PH + PI result), but to show (1) thatFiltSketch works with filtrations other than the Flood complex filtration and (2) that one can indeed almost match the performance of neural baselines with a topological approach.


W3: The computational advantage remains unverified. Despite efficiency being a central motivation, the paper reports no runtime, peak-memory, or empirical scaling comparisons. The complexity bound excludes filtration construction and involves thousands of filtration-level pairs. Measurements against optimized PH and spectral baselines are necessary to establish the practical accuracy–cost tradeoff.

We agree and refer the reviewer to our general comments section "Runtime and efficiency". In summary, we have now included a detailed runtime comparison (Appendix E) between FiltSketch and PH+PI, as well as a runtime breakdown of the different computational steps. Overall, our runtime study shows favorable scaling behavior of FiltSketch with smaller runtime than PH+PI when using the Flood complex and increasing the number of landmarks.

AuthorsOpenReview ↗

Response to Reviewer ttLC (2/2)

Q1: How do rank approximation error and clipping frequency vary with diffusion depth and probe count? How does an exact-rank stream perform with the same TCN?

We want to clarify that the Hodge and persistence rank features do not accurately approximate Betti numbers and persistence ranks at small kk. Instead, they encode the local structure of the complexes at the different filtration levels. For further details, see our response to reviewer av2404.

We also refer the reviewer to our general comment section "On persistent rank estimates vs. exact persistence ranks", where we also discuss an additional experiment that benchmarks the performance of exact persistence ranks (see Table 5, Appendix F). On the Rocks dataset, for example, balanced accuracy drops from 0.96 ± 0.01 to 0.93 ± 0.02 when switching to exact persistence ranks, while on QM7b we observe similar performance (MAE ratio of 0.72 vs. 0.73; lower is better).


Q2: Were the Davies et al. baselines recomputed using inverse-Coulomb filtrations? Are results available for PH baselines using the same directional cubical inputs?

Please see our response to your comment on the Davies et al. filtration construction.

We have now also added a PH+PI baseline using the same cubical complex filtration as for FiltSketch where, for each filtration direction, we compute H0, H1 and H2 PI's and feed them through a ResNet-style CNN. As with FiltSketch, aggregation is done via a DeepSet module and we use the same MLP classification head. We updated Table 2 with the corresponding results, which show that the PH+PI approach also benefits from the cubical filtration, yet classification balanced accuracy remains below that of our FiltSketch approach (84% vs. 92%).


Q3: What are the feature-extraction time, total preprocessing time, and peak memory relative to the principal baselines?

We kindly refer the reviewer to our general comments section "Runtime and efficiency".

Reviewer_av24OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes FiltSketch, a method for learning topological and geometric information without explicitly computing persistence barcodes or the full spectrum of Hodge Laplacians. The authors construct features from different filtration levels using two types of information. The first is Hodge spectral information, which describes the structure at individual filtration levels. The second is approximate persistent ranks, which describe how topological features survive between different levels. They estimate these quantities using sparse matrix operations and randomized trace estimation. The resulting features are processed by a temporal convolutional network for downstream prediction. The authors evaluate the approach on molecular property prediction and several point-cloud benchmarks. Their results show improvements over some topology-based baselines and competitive performance with point-cloud neural networks in certain settings.

Critical strengths
  • The paper presents an interesting approach to extracting topological information without explicitly computing persistence barcodes or performing eigendecomposition. The mathematical formulation, particularly the approximation of persistent ranks using sparse matrix operations, is interesting. The authors also provide derivations to support their method.
  • The experiments cover different problems, including molecular property prediction and point-cloud classification. The authors also include ablation experiments to investigate the effects of different features.
Critical weaknesses
  • I did not fully understand the computational advantage of using this method. The authors motivate their approach by explaining that computing persistence barcodes and spectral information is expensive. However, I could not find actual runtime or memory comparisons with existing methods. For example, the method still needs to construct the filtration and perform many matrix operations across different filtration levels. In the point-cloud experiments, they use 512 filtration levels and multiple random probes. Without reporting the actual computational cost, it is hard to understand whether avoiding explicit barcode computation provides a meaningful advantage. Also, the authors approximate persistent ranks, but do not show how close these approximations are to the exact values.
  • Another concern is that the experiments do not clearly separate the contributions of different components. For example, on ModelNet10, FiltSketch achieves only 74% balanced accuracy using the Flood filtration. After changing to cubical filtrations with multiple directional views, accuracy increases to 92%. However, the authors does not evaluate the other persistent-homology baselines with the same cubical filtrations. Therefore, it is unclear whether the improvement comes from FiltSketch itself or just using a better filtration. Also, the ablation experiments show that persistent-rank features alone perform close to the full method. For example, on Rocks, adding Hodge features improves accuracy marginally. I am not sure whether this marginal improvement justifies the additional complexity. The downstream network also contains several components, but there are limited ablation experiments showing whether all of them are necessary.
Questions for authors
  • Can the authors provide actual runtime and memory comparisons with existing persistent-homology methods? It would be useful to include filtration construction and feature computation.
  • How accurate are the approximate persistent ranks? Can the authors show how the approximation error changes when they increase the number of probes or diffusion depths?
  • Why are the existing persistent-homology baselines not evaluated using the same cubical filtrations on ModelNet10? Would the improvement remain similar under this comparison?
  • Can the authors show the performance using a simpler downstream model? It is unclear whether all components of the proposed architecture are necessary.
2 comments
AuthorsOpenReview ↗

Response to Reviewer av24 (1/2)

➤ W1: Computational cost and persistent rank approximation

I did not fully understand the computational advantage of using this method ... Without reporting the actual computational cost, it is hard to understand whether avoiding explicit barcode computation provides a meaningful advantage. Also, the authors approximate persistent ranks, but do not show how close these approximations are to the exact values.

We address both concerns in our general comment to all reviewers.

In summary, we have now added the requested runtime experiments to Appendix E of the updated PDF. These experiments clearly show an advantage (of up to 5x at 8k landmarks when using the Flood complex) against exact PH computation (again using the Flood complex) combined with persistence image (PI) construction (we chose PI's as they can be computed efficiently and thus present a strong runtime baseline).

We also recognize that the wording "approximate persistence ranks" was unfortunate, as our goal lies more in obtaining informative and computationally tractable features, rather than in achieving an accurate approximation. We have updated the manuscript in this regard and now also present experiments with exact persistence ranks for comparison (see Table 5).

➤ Contributions of different components ("Another concern is that the experiments do not clearly separate the contributions of different components.")

The authors does not evaluate the other persistent-homology baselines with the same cubical filtrations. Therefore, it is unclear whether the improvement comes from FiltSketch itself or just using a better filtration.

As requested, we have now added a PH+PI baseline using the same cubical complex filtration as for FiltSketch where, for each filtration direction, we compute H0, H1 and H2 PI's and feed them through a ResNet-style CNN. As with FiltSketch, aggregation is done via a DeepSet module and we use the same MLP classification head.
We have added the corresponding result to the updated Table 2, which shows that the PH+PI approach also benefits from the cubical filtration, yet classification balanced accuracy remains below that of our FiltSketch approach (84% vs. 92%).

We also want to clarify that we included the cubical complex filtration not to demonstrate an improvement of FiltSketch over other PH methods (which benefit as well from this construction, as can be seen from the added PH + PI result), but to show (1) thatFiltSketch works with filtrations other than the Flood complex filtration and (2) that one can indeed almost match the performance of neural baselines with a topological approach.

Also, the ablation experiments show that persistent-rank features alone perform close to the full method. For example, on Rocks, adding Hodge features improves accuracy marginally. I am not sure whether this marginal improvement justifies the additional complexity.

We agree that the Hodge stream only provides small improvements in our ablations, yet it does so consistently (also for the other datasets). Hence, we think it supplies valuable information to the learning task(s). Moreover, we want to refer to the newly added breakdown of the runtime in Figure 2, which shows that only a small part of the runtime is required exclusively for the Hodge stream.

The downstream network also contains several components, but there are limited ablation experiments showing whether all of them are necessary.

As requested, we added an ablation experiment for the downstream TCN:

FiltSketchBal. acc. ↑#Param.
+ full (7-layer) TCN, original readout0.97± 0.011,588,235
+ 1-layer TCN (ks 3), original readout0.93±0.02395,144
+ 1-layer TCN (ks 3), mean + linear readout0.82 ± 0.06261,639
+ 1-layer TCN (ks 1), mean + linear readout0.73 ± 0.07261,383

Overall, we can draw several conclusions from these results: First, the 1-layer TCN retains strong performance (0.93 versus 0.97 balanced accuracy) with approximately 75% fewer parameters, showin that the sketches remain predictive with a substantially shallower backbone. Second, replacing the original readout and nonlinear prediction head with mean pooling and a linear classifier reduces accuracy to 0.82, indicating that aggregation and prediction contribute considerably. Third, restricting the remaining convolution to kernel size one further reduces accuracy to 0.73 while removing only 256 parameters, supporting the value of local cross-timepoint interactions rather than parameter count alone.
In summary, the results demonstrate that useful predictive information is in fact accessible to simpler models, while also showing that the full downstream architecture improves its extraction.

The Appendix has been extended and adjusted accordingly (Table 7, Appendix F).

AuthorsOpenReview ↗

Response to Reviewer av24 (2/2)

Questions

Q1: Can the authors provide actual runtime and memory comparisons with existing persistent-homology methods? It would be useful to include filtration construction and feature computation.

As written above, we added runtime comparisons between FiltSketch and PH + PI computations (as both are the representations that are fed into the machine learning models). We deliberately did not include filtration construction times, as they are shared between both approaches. In practice, filtrations are represented as GUDHI SimplexTrees, and both PH and FiltSketch are computed on top.


Q2: How accurate are the approximate persistent ranks? Can the authors show how the approximation error changes when they increase the number of probes or diffusion depths?

The trace estimators are pretty precise; we address this concern in our general comment section "Accuracy of trace estimators".

However, the finite depth approximations are inherently biased. Calling them "spectral persistence summaries" would be more appropriate, and we already updated the manuscript in this regard.
In fact, the Hodge and persistence rank features do not accurately approximate Betti numbers and persistence ranks at small kk. Instead, they encode the local structure of the complexes at the different filtration levels.

To be more precise, the trace of (I−αL‾)k(I-\alpha \overline{L})^k depends only on closed walks of at most kk steps, so complexes that agree locally (in their ⌊k/2⌋\lfloor k/2 \rfloor-hop neighborhoods) cannot be told apart. Moreover, small positive eigenvalues λ≲1αk\lambda \lesssim \tfrac 1 {\alpha k} still contribute non-negligible amounts to the trace of (I−αL‾)k(I-\alpha \overline{L})^k. Thus, almost-harmonic chains are also counted as kernel elements at small kk, resulting in an overestimate of the kernel dimensions, while ranks are underestimated. Consequently, the Hodge and persistence rank features do not accurately approximate Betti numbers and persistence ranks at small kk. Instead, they encode the local structure of the complexes at the different filtration levels.

We also refer the reviewer to our general comment section "On persistent rank estimates vs. exact persistence ranks", where we also discuss an additional experiment that benchmarks the performance of exact persistence ranks (see Table 5, Appendix F).


Q3: Why are the existing persistent-homology baselines not evaluated using the same cubical filtrations on ModelNet10? Would the improvement remain similar under this comparison?

Please see our response to your earlier comment on missing baselines for the cubical filtrations on ModelNet10.


Q4: Can the authors show the performance using a simpler downstream model? It is unclear whether all components of the proposed architecture are necessary.

Please see our earlier results table to your comment on the performance of a simpler TCN model.

In addition to that, we ablate the effect of different downstream architectures on top of the FiltSketch and PI representations for the MCB-C classification balanced accuracy in the following table (now Table 6, see Appendix F for additional implementation details).

ModelPIFiltSketch
CNN(2D/1D)66.2±2.769.0±3.5
MLP60.9±5.770.5±3.5
kNN55.8±3.958.1±4.1
Random forest64.8±2.065.0±4.1
LightGBM69.5±3.867.6±4.1
XGBoost67.7±2.567.7±4.6
GRU--69.7±3.5
LSTM--63.4±4.2
TCN--77.3±3.0

We see two effects:

  • First, the standard machine learning methods achieve around the same balanced accuracy on both representations, but the neural networks (MLP and CNN) are able to **leverage more information from FiltSketch than from the PIs, as they achieve better classification performances (69.0% vs 66.2% for CNN and 60.9% vs 70.5% for MLP), which suggests that performance benefits to some part can be attributed to the FiltSketch representations.
  • Second, parts of the performance gains of FiltSketch over PI are due to the classification model itself, or rather to how the classification model leverages the sequential information within the filtration sketches. The GRU model achieves similar performance to the MLP (69.7% vs. 70.5%), and the LSTM performs clearly worse, while the TCN achieves by far the best performance across all models (77.3%).

Comments on the paper

2 comments
AuthorsOpenReview ↗

General comment to all Reviewers

We thank all reviewers for their very valuable comments and overall positive sentiment regarding the novelty and execution of our approach. We have revised the manuscript to address the reviewers’ feedback. Below, we first address questions regarding runtime and rank estimation that were raised by multiple reviewers, before turning to the remaining questions in our individual responses to each reviewer.


Runtime and efficiency

➤ Representation construction

We agree with reviewers that complexity bounds (as in the appendix) alone do not establish a practical computational advantage.

We have therefore added a runtime benchmark for representation computation on ten randomly selected Rocks examples, varying the Flood complex resolution from 1k to 8k landmarks. FiltSketch uses 512 sampling points, diffusion powers 1,2,4,8\\{1,2,4,8\\}, lags 1,4,16,64,256\\{1,4,16,64,256\\}, and eight trace probes. The PH baseline uses GUDHI's PH implementation followed by persistence images on a 128×128 grid. Each measurement runs 20 timed repetitions. Median representation-extraction times are comparable at 1k landmarks (37.2 versus 39.9 ms), while the median speedup increases to 4.84x at 8k landmarks (88.2 versus 443.2 ms). These measurements demonstrate favorable empirical scaling. Runtime measurements in both cases start from an already available GUDHI SimplexTree data structure; computation time for the latter is not included in the reported results, as the filtration needs to be built for FiltSketch and Flood PH + PI.

We have added a runtime section that matches this description in Appendix E (including a runtime breakdown).

Regarding memory requirements (in PyTorch), PH+PI used approx. 93 MiB of peak GPU allocation across all tested sizes due to chunked GPU rasterization during PI construction, while median peak CPU process memory increased from 0.73 to 0.80 GiB. FiltSketch used approx. twice the CPU process memory (1.33–1.40 GiB), with median peak GPU allocation increasing approx. linearly from 41 to 397 MiB as the number of Flood landmarks increases from 1k to 8k. Neither method encountered memory limitations in these experiments.

Flood landmarksFiltSketch GPU memoryPH+PI GPU memoryFiltSketch CPU RAMPH+PI CPU RAM
1k41 MiB93 MiB1.33 GiB0.73 GiB
2k87 MiB93 MiB1.35 GiB0.74 GiB
4k187 MiB93 MiB1.36 GiB0.76 GiB
8k397 MiB93 MiB1.40 GiB0.80 GiB

Rank estimation and trace estimation accuracy

➤ On persistent rank estimates vs. exact persistence ranks

We are aware that our approximate persistence ranks wording suggests closeness to the limiting values (i.e., the actual persistence ranks) that can break down at small diffusion depths. In fact, an accurate approximation is neither our goal, nor required for the validity of the method. Instead, we aim to extract informative geometric features of simplicial complexes that are not accessible from persistent homology alone. We will therefore rename these features to spectral persistence summaries and clarify that our objective is to retain task-relevant information, rather than minimize persistence-rank estimation error.

In other words, while the limits (wrt. the number of diffusion steps) of our computed quantities have a clear interpretation as Betti numbers and persistence ranks, we are much more interested in the positive-spectrum contributions retained in finite-depth traces and how they behave over increasing diffusion depths.

As requested, we have added an experiment (Table 5) where we replace the rank stream with exact persistence ranks. On the Rocks dataset balanced accuracy drops from 0.96 ± 0.01 to 0.93 ± 0.02, while on QM7b, both variants yield similar performance (MAE ratio of 0.72 vs. 0.73).

Overall, we think our current results support the usefulness of the spectral summaries. We, however, acknowledge that the variants differ in channel count (for the downstream TCN model) and coefficient field, and that size/degree effects may also contribute.

➤ Accuracy of trace estimators

To distinguish stochastic trace-estimation error from finite-depth approximation bias, we compare randomized estimates with exact finite-depth traces (not exact PH ranks) on five Rocks filtrations, using 7 filtration levels, diffusion powers 1,2,4,8\\{1,2,4,8\\} and 50 trials per probe count and report median errors below.

ProbesLower HodgeUpper HodgeFull Hodge
20.0050.0070.004
40.0030.0050.003
60.0030.0030.002
80.0020.0030.002
160.0010.0020.001

We can see that errors decrease consistently as the number of probes increases. Notably, the improvement from two to eight probes is close to the expected R−12R^{-\tfrac12} behavior.

AuthorsOpenReview ↗

Feedback before discussion period is ending

Dear Reviewers,

We would like to thank you all once more for your time and helpful comments and suggestions. We think that the requested additional experiments and clarifications have substantially improved the manuscript.
As the discussion phase ends tomorrow, we would like to kindly ask if your comments and questions have been addressed in our general comment (to all reviewers) and our per-reviewer responses. If so, we would be grateful if you could update your scores accordingly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.