pith. sign in

arxiv: 2604.22676 · v2 · submitted 2026-04-24 · 💻 cs.LG

Operational Feature Fingerprints of Graph Datasets via a White-Box Signal-Subspace Probe

Yuchen Xiong , Swee Keong Yeap , Zhen Hong Ban This is my paper

Pith reviewed 2026-05-08 12:06 UTC · model grok-4.3

classification 💻 cs.LG
keywords graph neural networkswhite-box interpretabilitydataset fingerprintingsignal subspacenode classificationridge fusion
0
0 comments X

The pith

A fixed dictionary of raw features, low-pass smoothing, and high-pass differences plus class-wise subspaces lets a white-box model match GNN accuracy while exposing which signals each graph dataset actually needs.

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

Graph neural networks hide the reasons for their node classifications inside entangled learned message passing. WG-SRC replaces that opacity with an explicit, named collection of graph operations: raw node features, row- and symmetric-normalized low-pass propagations, and high-pass differences. These signals are reduced by Fisher selection and class-specific PCA, then combined through closed-form multi-alpha ridge decisions whose weights are tuned on validation. The resulting predictor stays competitive with reproduced black-box baselines across six datasets, so its internal decomposition can be read as a trustworthy fingerprint of the dominant mechanisms required by each graph.

Core claim

WG-SRC replaces learned message passing with a fixed, named graph-signal dictionary containing raw features, row- and symmetric-normalized low-pass propagation, and high-pass graph differences. It combines Fisher coordinate selection, class-wise PCA subspaces, closed-form multi-alpha ridge classification, and validation-based score fusion so that both prediction and analysis rest on explicit class subspaces, energy-controlled dimensions, and closed-form linear decisions.

What carries the argument

The WG-SRC probe, which decomposes graph data into a fixed dictionary of raw, low-pass, and high-pass signals, projects them into class-wise PCA subspaces, and fuses them with ridge regression to produce both predictions and explicit dataset fingerprints.

If this is right

  • The atlas decomposes node classification into raw-feature, low-pass, high-pass, class-geometric, and ridge-boundary contributions.
  • Aligned interventions identify when high-pass blocks act as removable noise and when raw or graph-derived signals must be preserved.
  • Fingerprints distinguish low-pass-dominated Amazon graphs, mixed high-pass and class-geometrically complex Chameleon behavior, and raw- or boundary-sensitive WebKB graphs.
  • WG-SRC functions simultaneously as a competitive white-box classifier and as a probe for fingerprint-conditioned analysis of black-box model components.

Where Pith is reading between the lines

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

  • The fingerprints could guide selection of GNN architectures that emphasize the dominant signal type identified for a given dataset.
  • Similar fixed-dictionary probes could be adapted for link prediction or graph classification by extending the signal operations.
  • The approach supplies a concrete test for whether particular layers inside a black-box GNN are sensitive to the same raw-versus-low-pass-versus-high-pass distinctions.

Load-bearing premise

That the fixed, named graph-signal dictionary plus class-wise PCA subspaces and ridge fusion sufficiently capture the entangled mechanisms of learned message passing without major loss of expressive power.

What would settle it

Measure whether WG-SRC accuracy drops far below GNN baselines on a new dataset whose internal fingerprint claims to fully explain the required signals; a large unexplained performance gap would show the fixed dictionary misses essential interactions.

Figures

Figures reproduced from arXiv: 2604.22676 by Swee Keong Yeap, Yuchen Xiong, Zhen Hong Ban.

Figure 1
Figure 1. Figure 1: WG-SRC pipeline. The graph is converted into named signal blocks, discriminative coordinates are selected, and prediction is produced by fusing class-subspace residuals with a closed-form ridge boundary. simple baselines. WG-SRC follows the decoupling intuition but removes learned hidden layers: it constructs low-pass and high-pass signals explicitly and audits which ones are used. White-box and subspace l… view at source ↗
Figure 2
Figure 2. Figure 2: Split-level companion to Table view at source ↗
Figure 3
Figure 3. Figure 3: Main-text atlas evidence for operational dataset fingerprints. (a) The signal simplex summarizes family view at source ↗
Figure 4
Figure 4. Figure 4: Mechanism atlas as a dataset diagnostic probe. view at source ↗
Figure 5
Figure 5. Figure 5: Error signal shift. The lollipop value is the family-size-adjusted high-pass share of wrong nodes view at source ↗
Figure 6
Figure 6. Figure 6: Chameleon class-pair geometry and confusion constellation. The left panel shows pairwise PCA view at source ↗
read the original abstract

Graph neural networks achieve strong node-classification accuracy, but learned message passing entangles ego attributes, neighborhood smoothing, high-pass graph differences, class geometry, and classifier-boundary effects inside opaque representations. This obscures why nodes are classified as they are and which graph-learning mechanisms a dataset requires. We propose WG-SRC, a white-box signal-subspace probe for prediction and graph dataset diagnosis. WG-SRC replaces learned message passing with a fixed, named graph-signal dictionary containing raw features, row- and symmetric-normalized low-pass propagation, and high-pass graph differences. It combines Fisher coordinate selection, class-wise PCA subspaces, closed-form multi-alpha ridge classification, and validation-based score fusion, so prediction and analysis rely on explicit class subspaces, energy-controlled dimensions, and closed-form linear decisions. As a white-box graph-learning instrument, WG-SRC uses predictive performance to validate its diagnostics. Across six node-classification datasets, it remains competitive with reproduced baselines and achieves positive average gain under aligned splits. Its atlas decomposes behavior into raw-feature, low-pass, high-pass, class-geometric, and ridge-boundary components. The resulting fingerprints distinguish low-pass-dominated Amazon graphs, mixed high-pass and class-geometrically complex Chameleon behavior, and raw- or boundary-sensitive WebKB graphs. Aligned interventions show when high-pass blocks act as removable noise, when raw or graph-derived signals should be preserved, and when ridge correction matters. WG-SRC therefore serves both as a functioning white-box classifier and as a dataset-fingerprinting probe, enabling fingerprint-conditioned analysis of how black-box model components behave under different graph-signal conditions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes WG-SRC, a white-box signal-subspace probe for graph node classification. It replaces learned message passing with a fixed named dictionary of graph signals (raw features, row- and symmetric-normalized low-pass, high-pass differences), combined with Fisher selection, class-wise PCA subspaces, closed-form multi-alpha ridge classification, and validation-based score fusion. The method claims competitive accuracy with reproduced baselines and positive average gain across six datasets under aligned splits, while generating an atlas that decomposes performance into raw-feature, low-pass, high-pass, class-geometric, and ridge-boundary components to produce dataset fingerprints (e.g., low-pass-dominated Amazon graphs vs. mixed Chameleon behavior).

Significance. If the central claims hold, WG-SRC would provide a reproducible, closed-form instrument for diagnosing which graph-signal mechanisms a dataset requires, enabling fingerprint-conditioned analysis of black-box GNN components. The emphasis on explicit subspaces, energy-controlled dimensions, and validation fusion strengthens interpretability and reproducibility compared to opaque learned representations.

major comments (3)
  1. [§3] §3 (signal dictionary and class-wise PCA construction): the central claim that this fixed linear dictionary plus ridge fusion sufficiently approximates the entangled non-linear multi-hop mechanisms of learned message passing is load-bearing for both the white-box classifier and fingerprinting results. Standard GNNs include learnable transforms and non-linear activations that induce feature interactions absent from the named dictionary; without ablation against non-linear baselines or datasets known to require higher-order filters, the resulting atlas risks reflecting probe inductive bias rather than dataset signal requirements.
  2. [Experimental results] Experimental results (performance tables and average-gain claims): the abstract and results report competitive performance and positive average gain across six datasets but provide no error bars, exact baseline reproduction protocols, data exclusion rules, or statistical significance tests. This directly undermines the validation step that is used to certify the diagnostics and fingerprints.
  3. [§5] §5 (aligned interventions and fingerprint analysis): the claims that high-pass blocks act as removable noise or that ridge correction matters on specific graphs rest on the untested assumption that the probe's decomposition faithfully isolates the relevant mechanisms; quantitative measures tying each intervention back to the original GNN decision boundaries are needed to support the downstream analysis of black-box components.
minor comments (2)
  1. [Method] Notation for multi-alpha ridge fusion and validation score combination should be formalized with explicit equations to improve reproducibility.
  2. [Figures] Figure captions for the atlas and fingerprints should include quantitative thresholds used for component dominance labeling.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the scope and limitations of WG-SRC as a diagnostic probe. We respond point by point below, indicating revisions where appropriate.

read point-by-point responses

  1. Referee: [§3] §3 (signal dictionary and class-wise PCA construction): the central claim that this fixed linear dictionary plus ridge fusion sufficiently approximates the entangled non-linear multi-hop mechanisms of learned message passing is load-bearing for both the white-box classifier and fingerprinting results. Standard GNNs include learnable transforms and non-linear activations that induce feature interactions absent from the named dictionary; without ablation against non-linear baselines or datasets known to require higher-order filters, the resulting atlas risks reflecting probe inductive bias rather than dataset signal requirements.

    Authors: We agree the fixed linear dictionary cannot fully replicate non-linear feature interactions or higher-order filters in GNNs. WG-SRC is positioned as an interpretable probe that isolates named signal types (raw, low-pass, high-pass) rather than as a complete surrogate for learned message passing. Competitive accuracy against reproduced non-linear GNN baselines on the six datasets indicates these components capture substantial predictive signal, and the atlas is validated by performance-based decomposition rather than exact replication. We will revise §3 to explicitly state this scope limitation, add discussion of potential inductive bias on datasets requiring complex non-linearities, and include a brief ablation note referencing performance on Chameleon (mixed high-pass behavior) to illustrate where the probe remains informative despite the linear constraint. revision: partial

  2. Referee: [Experimental results] Experimental results (performance tables and average-gain claims): the abstract and results report competitive performance and positive average gain across six datasets but provide no error bars, exact baseline reproduction protocols, data exclusion rules, or statistical significance tests. This directly undermines the validation step that is used to certify the diagnostics and fingerprints.

    Authors: We accept this criticism. The current version lacks error bars, detailed reproduction protocols, and significance testing, which weakens the reported gains. In revision we will: (i) add standard error bars from multiple random seeds, (ii) expand the experimental setup with exact baseline hyperparameters, split alignments, and data exclusion criteria, and (iii) include paired statistical tests (e.g., t-tests) on the average-gain claims. These additions will directly support the validation of the fingerprint diagnostics. revision: yes

  3. Referee: [§5] §5 (aligned interventions and fingerprint analysis): the claims that high-pass blocks act as removable noise or that ridge correction matters on specific graphs rest on the untested assumption that the probe's decomposition faithfully isolates the relevant mechanisms; quantitative measures tying each intervention back to the original GNN decision boundaries are needed to support the downstream analysis of black-box components.

    Authors: The interventions are validated through measurable accuracy changes when specific signal blocks are removed or adjusted, providing empirical support for the decomposition. However, we acknowledge that direct quantitative linkage to black-box GNN decision boundaries (e.g., via attribution alignment) is not performed. We will revise §5 to clarify that the probe offers correlative rather than causal isolation, add a limitations paragraph noting this gap, and include a new quantitative measure: correlation between component-wise performance deltas and GNN attribution scores (computed via integrated gradients on a subset of models) to better tie interventions to original boundaries. revision: partial

Circularity Check

0 steps flagged

No significant circularity detected; derivation remains self-contained

full rationale

The WG-SRC construction uses an explicitly enumerated fixed dictionary of graph signals (raw features, normalized low-pass and high-pass differences), class-wise PCA subspaces, closed-form multi-alpha ridge regression, and validation-driven score fusion. None of these steps reduce a claimed prediction or fingerprint to a quantity defined by the same fitted parameters or by a self-citation chain; the closed-form solutions and external validation splits keep the pipeline independent of the target outputs. No uniqueness theorems, ansatzes smuggled via prior work, or renamings of known results appear as load-bearing elements in the provided derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that a small fixed dictionary of graph signals plus linear subspace methods can serve as a faithful white-box proxy for learned GNN message passing; no explicit free parameters are named in the abstract but ridge alphas and fusion weights are implied.

axioms (1)
  • domain assumption The fixed named graph-signal dictionary (raw features, normalized low-pass, high-pass differences) captures the essential mechanisms of graph learning
    Invoked when replacing learned message passing with the dictionary for both prediction and diagnosis.

pith-pipeline@v0.9.0 · 5604 in / 1404 out tokens · 45067 ms · 2026-05-08T12:06:32.770210+00:00 · methodology

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

24 extracted references · 2 canonical work pages

  1. [1]

    Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks. International Conference on Learning Representations, 2017

    2017

  2. [2]

    L., Ying, Z., and Leskovec, J

    Hamilton, W. L., Ying, Z., and Leskovec, J. Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems 30 (NeurIPS 2017), pp. 1024--1034, 2017

    2017

  3. [3]

    Graph attention networks

    Veli c kovi \'c , P., Cucurull, G., Casanova, A., Romero, A., Li \`o , P., and Bengio, Y. Graph attention networks. International Conference on Learning Representations, 2018

    2018

  4. [4]

    Predict then propagate: Graph neural networks meet personalized PageRank

    Gasteiger, J., Bojchevski, A., and G \"u nnemann, S. Predict then propagate: Graph neural networks meet personalized PageRank. International Conference on Learning Representations, 2019

    2019

  5. [5]

    Huang, Q., He, H., Singh, A., Lim, S.-N., and Benson, A. R. Combining label propagation and simple models out-performs graph neural networks. International Conference on Learning Representations, 2021

    2021

  6. [6]

    Simple and deep graph convolutional networks

    Chen, M., Wei, Z., Huang, Z., Ding, B., and Li, Y. Simple and deep graph convolutional networks. In Proceedings of the 37th International Conference on Machine Learning, PMLR 119:1725--1735, 2020

    2020

  7. [7]

    Sign: Scalable inception graph neural networks

    Frasca, F., Rossi, E., Eynard, D., Chamberlain, B., Bronstein, M. M., and Monti, F. SIGN: Scalable inception graph neural networks. arXiv preprint arXiv:2004.11198, 2020

    work page arXiv 2004

  8. [8]

    C.-C., Lei, Y., and Yang, B

    Pei, H., Wei, B., Chang, K. C.-C., Lei, Y., and Yang, B. Geom-GCN: Geometric graph convolutional networks. International Conference on Learning Representations, 2020

    2020

  9. [9]

    Beyond homophily in graph neural networks: Current limitations and effective designs

    Zhu, J., Yan, Y., Zhao, L., Heimann, M., Akoglu, L., and Koutra, D. Beyond homophily in graph neural networks: Current limitations and effective designs. Advances in Neural Information Processing Systems, 2020

    2020

  10. [10]

    L., Gupta, V., Bhalerao, O., and Lim, S.-N

    Lim, D., Hohne, F., Li, X., Huang, S. L., Gupta, V., Bhalerao, O., and Lim, S.-N. Large scale learning on non-homophilous graphs: New benchmarks and strong simple methods. Advances in Neural Information Processing Systems, 2021

    2021

  11. [11]

    Adaptive universal generalized PageRank graph neural network

    Chien, E., Peng, J., Li, P., and Milenkovic, O. Adaptive universal generalized PageRank graph neural network. International Conference on Learning Representations, 2021

    2021

  12. [12]

    and Lenssen, J

    Fey, M. and Lenssen, J. E. Fast graph representation learning with PyTorch Geometric. ICLR Workshop on Representation Learning on Graphs and Manifolds, 2019

    2019

  13. [13]

    Multi-scale attributed node embedding

    Rozemberczki, B., Allen, C., and Sarkar, R. Multi-scale attributed node embedding. Journal of Complex Networks, 9(2):cnab014, 2021

    2021

  14. [14]

    Yu, Y., Chan, K. H. R., You, C., Song, C., and Ma, Y. Learning diverse and discriminative representations via the principle of maximal coding rate reduction. Advances in Neural Information Processing Systems, 2020

    2020

  15. [15]

    Chan, K. H. R., Yu, Y., You, C., Qi, H., Wright, J., and Ma, Y. ReduNet: A white-box deep network from the principle of maximizing rate reduction. Journal of Machine Learning Research, 23(114):1--103, 2022

    2022

  16. [16]

    A global geometric analysis of maximal coding rate reduction

    Wang, P., Liu, H., Pai, D., Yu, Y., Zhu, Z., Qu, Q., and Ma, Y. A global geometric analysis of maximal coding rate reduction. International Conference on Machine Learning, 2024

    2024

  17. [17]

    On lines and planes of closest fit to systems of points in space

    Pearson, K. On lines and planes of closest fit to systems of points in space. Philosophical Magazine, 2(11):559--572, 1901

    1901

  18. [18]

    Hoerl, A. E. and Kennard, R. W. Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12(1):55--67, 1970

    1970

  19. [19]

    A tutorial on spectral clustering

    von Luxburg, U. A tutorial on spectral clustering. Statistics and Computing, 17(4):395--416, 2007

    2007

  20. [20]

    Pitfalls of graph neural network evaluation

    Shchur, O., Mumme, M., Bojchevski, A., and G \"u nnemann, S. Pitfalls of graph neural network evaluation. Relational Representation Learning Workshop, NeurIPS, 2018

    2018

  21. [21]

    SGFormer: Simplifying and empowering transformers for large-graph representations

    Wu, Q., Zhao, W., Yang, C., Zhang, H., Nie, F., Jiang, H., Bian, Y., and Yan, J. SGFormer: Simplifying and empowering transformers for large-graph representations. Advances in Neural Information Processing Systems 36, 2023

    2023

  22. [22]

    B., and Goldstein, T

    Kong, K., Chen, J., Kirchenbauer, J., Ni, R., Bruss, C. B., and Goldstein, T. GOAT: A global transformer on large-scale graphs. In Proceedings of the 40th International Conference on Machine Learning, pp. 17375--17390, 2023

    2023

  23. [23]

    NodeFormer: A scalable graph structure learning transformer for node classification

    Wu, Q., Zhao, W., Li, Z., Wipf, D., and Yan, J. NodeFormer: A scalable graph structure learning transformer for node classification. In Advances in Neural Information Processing Systems, 2022

    2022

  24. [24]

    Mixture of experts meets decoupled message passing: Towards general and adaptive node classification

    Chen, X., Zhou, J., Yu, S., and Xuan, Q. Mixture of experts meets decoupled message passing: Towards general and adaptive node classification. arXiv preprint arXiv:2412.08193, 2024

    work page arXiv 2024