Sampling Transferable Graph Neural Networks with Limited Graph Information

Gonzalo Mateos; Haoyu Wang; Luana Ruiz; Renyuan Ma

arxiv: 2410.16593 · v6 · submitted 2024-10-22 · 📡 eess.SP · cs.AI· cs.LG

Sampling Transferable Graph Neural Networks with Limited Graph Information

Haoyu Wang , Renyuan Ma , Gonzalo Mateos , Luana Ruiz This is my paper

Pith reviewed 2026-05-23 19:25 UTC · model grok-4.3

classification 📡 eess.SP cs.AIcs.LG

keywords graph neural networkssubgraph samplingtransferabilityLaplacian tracefeature alignmentspectral propertieshomophily

0 comments

The pith

Feature-driven sampling that maximizes the Laplacian trace produces GNNs with better transferability from limited graph information.

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

The paper establishes that when precise graph connectivity is unavailable, subgraph sampling guided by node features can still preserve spectral properties that control GNN expressivity and transferability. It introduces two alignment notions: a coarse measure of feature homophily that lower-bounds the Laplacian trace, and a fine stationary model linking feature covariance diagonals to node degrees under monotone filters. These motivate a non-sequential sampling algorithm on the feature matrix alone. On real-world benchmarks the retention rule maximizing the trace proxy consistently yields superior transferability and smaller generalization gaps.

Core claim

Selecting the retention rule that maximizes the Laplacian trace in feature-driven sampling consistently yields GNNs with superior transferability and reduced generalization gaps on real-world benchmarks.

What carries the argument

Alignment-based perspective linking node feature statistics to graph spectral structure, using the Laplacian trace as a proxy for operator rank preserved by sampling.

If this is right

Coarse alignment supplies a lower bound on Laplacian trace directly from feature statistics.
Fine alignment shows that filter monotonicity dictates how feature variance relates to spectral energy.
The sampling algorithm operates directly on the feature matrix without needing sequential graph access.
Trace-maximizing rules reduce generalization gaps when training on sampled subgraphs and testing on larger target graphs.

Where Pith is reading between the lines

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

The approach could apply to settings where only node attributes are observed and edges must be inferred from features.
Controlled synthetic experiments varying the degree of eigenbasis alignment would isolate how much the trace proxy drives the gains.
Combining the trace criterion with existing degree or random sampling heuristics might produce hybrid rules that are robust when the stationary assumption is only approximate.

Load-bearing premise

Feature covariance and the Laplacian share an eigenbasis under the stationary model so that covariance diagonals reflect node-degree ordering.

What would settle it

Finding that a non-trace-maximizing retention rule produces equal or better transferability and smaller gaps on the same real-world benchmarks would falsify the central empirical claim.

Figures

Figures reproduced from arXiv: 2410.16593 by Gonzalo Mateos, Haoyu Wang, Luana Ruiz, Renyuan Ma.

**Figure 1.** Figure 1: Adjusted Laplacian trace versus graph subsampling rate. The adjusted trace is the subsampled graph Laplacian trace normalized by the number of [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Figure 2: Example of randomly sampled subgraph (green) and subgraph sampled [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Test accuracy achieved by GNN on full graph versus training graph subsampling rate. Boxplots indicate the spread of the accuracy realized by GNNs [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

read the original abstract

Graph neural networks (GNNs) achieve strong performance on graph learning tasks, but training on large-scale networks remains computationally challenging. Transferability results show that GNNs with fixed weights can generalize from smaller graphs to larger ones drawn from the same family, motivating the use of sampled subgraphs to boost training efficiency. Yet most existing sampling strategies rely on reliable access to the target graph structure, which in practice may be noisy, incomplete, or unavailable prior to training. In lieu of precise connectivity information, we study feature-driven subgraph sampling for transferable GNNs, with the goal of preserving spectral properties of graph operators that control GNN expressivity. We adopt an alignment-based perspective linking node feature statistics to graph spectral structure and develop two complementary notions of feature-graph alignment. For coarse alignment, we formalize feature homophily through a Laplacian-based measure quantifying the alignment of feature principal components with graph eigenvectors, and establish a lower bound on the Laplacian trace in terms of feature statistics. This motivates a simple, non-sequential sampling algorithm that operates on the feature matrix and preserves a trace-based proxy for operator rank. For fine alignment, we assume a stationary model where the feature covariance and Laplacian share an eigenbasis, and establish that diagonal covariance entries reflect node-degree ordering under monotone filters. We empirically validate that filter monotonicity dictates the relationship between feature variance and spectral energy. On real-world benchmarks, selecting the retention rule that maximizes the Laplacian trace consistently yields GNNs with superior transferability and reduced generalization gaps.

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.

Desk Editor's Note private letter to a colleague

Feature-driven sampling for transferable GNNs is a sensible direction but the headline empirical result requires full target-graph access to pick the retention rule, which undercuts the limited-information premise.

read the letter

The paper's main move is to replace structure-dependent sampling with a feature-matrix approach that tries to preserve spectral properties relevant to GNN transferability. It introduces two alignment notions: a coarse one that lower-bounds the Laplacian trace using feature principal components, and a fine one that assumes feature covariance and the Laplacian share an eigenbasis so that variance ordering reflects degree under monotone filters. The sampling algorithm itself is non-sequential and runs only on the feature matrix, which is the practical part worth noting. On the positive side, the trace-based proxy is a clean way to connect feature homophily to operator rank, and the empirical section does show reduced generalization gaps when the chosen rule happens to maximize the trace. That part is straightforward and could be useful to people already working on scalable training. The soft spot is the selection step. The abstract states that picking the retention rule which maximizes the Laplacian trace gives the best transferability on benchmarks. Computing that trace needs the true Laplacian of the target graph. The whole point of the work is sampling when that graph is noisy or unavailable, so the reported procedure cannot be executed in the setting the title and abstract advertise. The stationary eigenbasis assumption is also strong and would need more justification or sensitivity checks, but the selection inconsistency is the load-bearing issue. This work is aimed at signal-processing and network-analysis groups that train GNNs on incomplete large graphs. A reader already thinking about feature-only proxies might pick up the alignment framing, but anyone expecting a method that truly operates without reliable connectivity will hit the gap. I would send it for review only if the authors can demonstrate a selection rule that stays within the feature-only constraint; otherwise the central claim does not hold up as stated.

Referee Report

2 major / 1 minor

Summary. The paper claims that feature-driven subgraph sampling, based on coarse and fine notions of feature-graph alignment, enables training of transferable GNNs when only limited or noisy graph structure is available. It derives a lower bound on the Laplacian trace from feature statistics to motivate a non-sequential sampling algorithm operating solely on the feature matrix, assumes a stationary model where feature covariance and Laplacian share an eigenbasis to link variance to node degrees under monotone filters, and reports that on real-world benchmarks, retention rules selected by maximizing the true Laplacian trace produce GNNs with better transferability and smaller generalization gaps.

Significance. If the central claims hold after addressing the applicability gap, the work would offer a practical route to scalable GNN training that avoids reliance on full target-graph connectivity, grounded in spectral alignment between features and operators. The trace lower bound and alignment measures could inform sampling design beyond the specific setting.

major comments (2)

[Abstract / empirical validation] Abstract and empirical validation: the headline result states that 'selecting the retention rule that maximizes the Laplacian trace consistently yields GNNs with superior transferability.' However, this selection step requires the true target Laplacian, which directly contradicts the problem premise of sampling with limited/noisy/unavailable graph structure. The coarse-alignment algorithm is feature-only, yet the reported procedure for choosing among retention rules cannot be executed under the stated constraints; this renders the empirical support for the central claim inapplicable to the motivating setting.
[Fine alignment paragraph] Fine alignment paragraph: the stationary-model assumption that feature covariance and Laplacian share an eigenbasis is used to conclude that diagonal covariance entries reflect node-degree ordering under monotone filters. No sensitivity analysis or counter-example is provided for graphs where this shared eigenbasis fails, yet the assumption underpins the claimed relationship between feature variance and spectral energy that is validated empirically.

minor comments (1)

[Coarse alignment section] Notation for the trace-based proxy and alignment measures should be introduced with explicit definitions before their use in the sampling algorithm description.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight important aspects of applicability and modeling assumptions. We address each major comment below and outline revisions to strengthen the manuscript.

read point-by-point responses

Referee: [Abstract / empirical validation] Abstract and empirical validation: the headline result states that 'selecting the retention rule that maximizes the Laplacian trace consistently yields GNNs with superior transferability.' However, this selection step requires the true target Laplacian, which directly contradicts the problem premise of sampling with limited/noisy/unavailable graph structure. The coarse-alignment algorithm is feature-only, yet the reported procedure for choosing among retention rules cannot be executed under the stated constraints; this renders the empirical support for the central claim inapplicable to the motivating setting.

Authors: We agree this is a valid concern: the reported experiments select retention rules by maximizing the true Laplacian trace, which is unavailable under the problem constraints. The coarse-alignment section derives a feature-only lower bound on the trace that serves as a practical proxy. In revision we will (i) explicitly distinguish oracle validation of the trace-maximization principle from the deployable feature-based selection rule, (ii) add experiments that rank retention rules using only the feature-derived lower bound and show comparable transferability gains, and (iii) update the abstract and empirical claims to reflect that the proxy is the operational method. These changes directly close the applicability gap while preserving the central result. revision: yes
Referee: [Fine alignment paragraph] Fine alignment paragraph: the stationary-model assumption that feature covariance and Laplacian share an eigenbasis is used to conclude that diagonal covariance entries reflect node-degree ordering under monotone filters. No sensitivity analysis or counter-example is provided for graphs where this shared eigenbasis fails, yet the assumption underpins the claimed relationship between feature variance and spectral energy that is validated empirically.

Authors: The stationary-model assumption is stated explicitly as a modeling device that yields the variance-to-degree relationship under monotone filters. Real-world empirical results are presented without claiming the assumption holds exactly on those graphs. To address the referee's point we will add a short robustness subsection that (a) quantifies eigenbasis alignment on the benchmark graphs and (b) includes a synthetic counter-example where the shared-eigenbasis condition is deliberately violated, showing how the variance-spectral-energy correlation degrades. This provides the requested sensitivity context without altering the core derivation. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The provided abstract and context describe explicit lower bounds on Laplacian trace from feature statistics, a stationary model assumption, and a feature-matrix sampling algorithm using a trace-based proxy. No quoted equations or steps reduce a claimed prediction to a fitted input by construction, invoke self-citations for load-bearing uniqueness, or smuggle ansatzes. The empirical selection of retention rules is presented as validation on benchmarks rather than a self-referential derivation. The derivation chain remains self-contained against external benchmarks with independent content.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claims rest on domain assumptions about feature-graph alignment and a stationary model; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (2)

domain assumption Feature covariance and Laplacian share an eigenbasis under a stationary model
Invoked for the fine alignment to link diagonal covariance entries to node-degree ordering.
domain assumption Monotone filters dictate the relationship between feature variance and spectral energy
Used to establish the empirical validation of filter monotonicity.

pith-pipeline@v0.9.0 · 5809 in / 1300 out tokens · 33112 ms · 2026-05-23T19:25:29.981532+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages

[1]

A new model for learning in graph domains,

Marco Gori, Gabriele Monfardini, and Franco Scarselli, “A new model for learning in graph domains,” in Proceedings. 2005 IEEE international joint conference on neural networks, 2005. IEEE, 2005, vol. 2, pp. 729– 734

work page 2005
[2]

Semi-supervised classification with graph convolutional networks,

T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in 5th Int. Conf. Learning Representations , Toulon, France, 24-26 Apr. 2017, Assoc. Comput. Linguistics

work page 2017
[3]

Convolutional neural networks on graphs with fast localized spectral filtering,

M. Defferrard, X. Bresson, and P. Vandergheynst, “Convolutional neural networks on graphs with fast localized spectral filtering,” in Neural Inform. Process. Syst. , Barcelona, Spain, 5-10 Dec. 2016, NIPS Foundation

work page 2016
[4]

Convolutional neural network architectures for signals supported on graphs,

F. Gama, A. G. Marques, G. Leus, and A. Ribeiro, “Convolutional neural network architectures for signals supported on graphs,” IEEE Trans. Signal Process., vol. 67, pp. 1034–1049, 2018

work page 2018
[5]

Graph neural networks: Architec- tures, stability and transferability,

L. Ruiz, F. Gama, and A. Ribeiro, “Graph neural networks: Architec- tures, stability and transferability,” Proc. IEEE , vol. 109, no. 5, pp. 660–682, 2021

work page 2021
[6]

Inductive repre- sentation learning on large graphs,

Will Hamilton, Zhitao Ying, and Jure Leskovec, “Inductive repre- sentation learning on large graphs,” Advances in neural information processing systems, vol. 30, 2017

work page 2017
[7]

Invariance- preserving localized activation functions for graph neural networks,

L. Ruiz, F. Gama, A. G. Marques, and A. Ribeiro, “Invariance- preserving localized activation functions for graph neural networks,” IEEE Trans. Signal Process. , vol. 68, pp. 127–141, 2020

work page 2020
[8]

Stability properties of graph neural networks,

F. Gama, J. Bruna, and A. Ribeiro, “Stability properties of graph neural networks,” IEEE Trans. Signal Process., vol. 68, pp. 5680–5695, 2020

work page 2020
[9]

Graphon neural networks and the transferability of graph neural networks,

L. Ruiz, L. F. O. Chamon, and A. Ribeiro, “Graphon neural networks and the transferability of graph neural networks,” in 34th Neural Inform. Process. Syst. , Vancouver, BC (Virtual), 6-12 Dec. 2020, NeurIPS Foundation

work page 2020
[10]

Trans- ferability of spectral graph convolutional neural networks,

R. Levie, W. Huang, L. Bucci, M. Bronstein, and G. Kutyniok, “Trans- ferability of spectral graph convolutional neural networks,” J. Mach. Learning Res., vol. 22, no. 272, pp. 1–59, 2021

work page 2021
[11]

Graphon signal processing,

L. Ruiz, L. F. O. Chamon, and A. Ribeiro, “Graphon signal processing,” IEEE Trans. Signal Process. , vol. 69, pp. 4961–4976, 2021

work page 2021
[12]

A spectral analysis of graph neural networks on dense and sparse graphs,

Luana Ruiz, Ningyuan Teresa Huang, and Soledad Villar, “A spectral analysis of graph neural networks on dense and sparse graphs,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 9936–9940

work page 2024
[13]

Discrete signal processing on graphs: Sampling theory,

S. Chen, R. Varma, A. Sandryhaila, and J. Kovacevic, “Discrete signal processing on graphs: Sampling theory,” IEEE Trans. Signal Process. , vol. 63, pp. 6510–6523, 2015

work page 2015
[14]

Efficient sampling set selection for bandlimited graph signals using graph spectral proxies,

Aamir Anis, Akshay Gadde, and Antonio Ortega, “Efficient sampling set selection for bandlimited graph signals using graph spectral proxies,” IEEE Transactions on Signal Processing, vol. 64, no. 14, pp. 3775–3789, 2016

work page 2016
[15]

A local graph limits perspective on sampling-based gnns,

Yeganeh Alimohammadi, Luana Ruiz, and Amin Saberi, “A local graph limits perspective on sampling-based gnns,” 2023

work page 2023
[16]

The emerging field of signal processing on graphs: Extending high- dimensional data analysis to networks and other irregular domains,

D. I Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst, “The emerging field of signal processing on graphs: Extending high- dimensional data analysis to networks and other irregular domains,” IEEE Signal Process. Mag. , vol. 30, no. 3, pp. 83–98, May 2013

work page 2013
[17]

Discrete signal processing on graphs,

A. Sandryhaila and J. M. F. Moura, “Discrete signal processing on graphs,” IEEE Trans. Signal Process. , vol. 61, pp. 1644–1656, Apr. 2013

work page 2013
[18]

Graph signal processing: Overview, challenges, and ap- plications,

A. Ortega, P. Frossard, J. Kova ˇcevi´c, J. M. F. Moura, and P. Van- dergheynst, “Graph signal processing: Overview, challenges, and ap- plications,” Proc. IEEE, vol. 106, no. 5, pp. 808–828, 2018

work page 2018
[19]

Optimal graph-filter design and applications to distributed linear network operators,

S. Segarra, A. G. Marques, and A. Ribeiro, “Optimal graph-filter design and applications to distributed linear network operators,” IEEE Trans. Signal Process., vol. 65, pp. 4117–4131, Aug. 2017

work page 2017
[20]

Convergent sequences of dense graphs i: Subgraph frequencies, metric properties and testing,

C. Borgs, J. T. Chayes, L. Lov ´asz, V . T. S ´os, and K. Vesztergombi, “Convergent sequences of dense graphs i: Subgraph frequencies, metric properties and testing,” Adv. Math. , vol. 219, no. 6, pp. 1801–1851, 2008

work page 2008
[21]

Lov ´asz, Large Networks and Graph Limits , vol

L. Lov ´asz, Large Networks and Graph Limits , vol. 60, American Mathematical Society, 2012

work page 2012
[22]

Transferability properties of graph neural networks,

L. Ruiz, L. F. O. Chamon, and A. Ribeiro, “Transferability properties of graph neural networks,” arXiv:2112.04629 [eess.SP]. Submitted to IEEE TSP, 2021

work page arXiv 2021
[23]

How powerful are graph neural networks?,

K. Xu, W. Hu, J. Leskovec, and S. Jegelka, “How powerful are graph neural networks?,” in 7th Int. Conf. Learning Representations , New Orleans, LA, 6-9 May 2019, pp. 1–17, Assoc. Comput. Linguistics

work page 2019
[24]

On the equivalence between graph isomorphism testing and function approxi- mation with gnns,

Zhengdao Chen, Soledad Villar, Lei Chen, and Joan Bruna, “On the equivalence between graph isomorphism testing and function approxi- mation with gnns,” Advances in neural information processing systems , vol. 32, 2019

work page 2019
[25]

How to learn a graph from smooth signals,

Vassilis Kalofolias, “How to learn a graph from smooth signals,” in Pro- ceedings of the 19th International Conference on Artificial Intelligence and Statistics (AISTATS). PMLR, 2016, pp. 920–929

work page 2016
[26]

Graph sparsification by effective resistances,

Daniel A. Spielman and Nikhil Srivastava, “Graph sparsification by effective resistances,” SIAM Journal on Computing , vol. 40, no. 6, pp. 1913–1926, 2011

work page 1913
[27]

ADAM: A method for stochastic optimization,

D. P. Kingma and J. L. Ba, “ADAM: A method for stochastic optimization,” in 3rd Int. Conf. Learning Representations , San Diego, CA, 7-9 May 2015, Assoc. Comput. Linguistics. APPENDIX Proof of Prop. II.2. From the definition, we can rewrite Y = span({Skx : k = 0, . . . , K}) Since S is symmetric, there exists some E, Σ ∈ Rn×n satis- fying EET = In and Σ d...

work page 2015

[1] [1]

A new model for learning in graph domains,

Marco Gori, Gabriele Monfardini, and Franco Scarselli, “A new model for learning in graph domains,” in Proceedings. 2005 IEEE international joint conference on neural networks, 2005. IEEE, 2005, vol. 2, pp. 729– 734

work page 2005

[2] [2]

Semi-supervised classification with graph convolutional networks,

T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in 5th Int. Conf. Learning Representations , Toulon, France, 24-26 Apr. 2017, Assoc. Comput. Linguistics

work page 2017

[3] [3]

Convolutional neural networks on graphs with fast localized spectral filtering,

M. Defferrard, X. Bresson, and P. Vandergheynst, “Convolutional neural networks on graphs with fast localized spectral filtering,” in Neural Inform. Process. Syst. , Barcelona, Spain, 5-10 Dec. 2016, NIPS Foundation

work page 2016

[4] [4]

Convolutional neural network architectures for signals supported on graphs,

F. Gama, A. G. Marques, G. Leus, and A. Ribeiro, “Convolutional neural network architectures for signals supported on graphs,” IEEE Trans. Signal Process., vol. 67, pp. 1034–1049, 2018

work page 2018

[5] [5]

Graph neural networks: Architec- tures, stability and transferability,

L. Ruiz, F. Gama, and A. Ribeiro, “Graph neural networks: Architec- tures, stability and transferability,” Proc. IEEE , vol. 109, no. 5, pp. 660–682, 2021

work page 2021

[6] [6]

Inductive repre- sentation learning on large graphs,

Will Hamilton, Zhitao Ying, and Jure Leskovec, “Inductive repre- sentation learning on large graphs,” Advances in neural information processing systems, vol. 30, 2017

work page 2017

[7] [7]

Invariance- preserving localized activation functions for graph neural networks,

L. Ruiz, F. Gama, A. G. Marques, and A. Ribeiro, “Invariance- preserving localized activation functions for graph neural networks,” IEEE Trans. Signal Process. , vol. 68, pp. 127–141, 2020

work page 2020

[8] [8]

Stability properties of graph neural networks,

F. Gama, J. Bruna, and A. Ribeiro, “Stability properties of graph neural networks,” IEEE Trans. Signal Process., vol. 68, pp. 5680–5695, 2020

work page 2020

[9] [9]

Graphon neural networks and the transferability of graph neural networks,

L. Ruiz, L. F. O. Chamon, and A. Ribeiro, “Graphon neural networks and the transferability of graph neural networks,” in 34th Neural Inform. Process. Syst. , Vancouver, BC (Virtual), 6-12 Dec. 2020, NeurIPS Foundation

work page 2020

[10] [10]

Trans- ferability of spectral graph convolutional neural networks,

R. Levie, W. Huang, L. Bucci, M. Bronstein, and G. Kutyniok, “Trans- ferability of spectral graph convolutional neural networks,” J. Mach. Learning Res., vol. 22, no. 272, pp. 1–59, 2021

work page 2021

[11] [11]

Graphon signal processing,

L. Ruiz, L. F. O. Chamon, and A. Ribeiro, “Graphon signal processing,” IEEE Trans. Signal Process. , vol. 69, pp. 4961–4976, 2021

work page 2021

[12] [12]

A spectral analysis of graph neural networks on dense and sparse graphs,

Luana Ruiz, Ningyuan Teresa Huang, and Soledad Villar, “A spectral analysis of graph neural networks on dense and sparse graphs,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 9936–9940

work page 2024

[13] [13]

Discrete signal processing on graphs: Sampling theory,

S. Chen, R. Varma, A. Sandryhaila, and J. Kovacevic, “Discrete signal processing on graphs: Sampling theory,” IEEE Trans. Signal Process. , vol. 63, pp. 6510–6523, 2015

work page 2015

[14] [14]

Efficient sampling set selection for bandlimited graph signals using graph spectral proxies,

Aamir Anis, Akshay Gadde, and Antonio Ortega, “Efficient sampling set selection for bandlimited graph signals using graph spectral proxies,” IEEE Transactions on Signal Processing, vol. 64, no. 14, pp. 3775–3789, 2016

work page 2016

[15] [15]

A local graph limits perspective on sampling-based gnns,

Yeganeh Alimohammadi, Luana Ruiz, and Amin Saberi, “A local graph limits perspective on sampling-based gnns,” 2023

work page 2023

[16] [16]

The emerging field of signal processing on graphs: Extending high- dimensional data analysis to networks and other irregular domains,

D. I Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst, “The emerging field of signal processing on graphs: Extending high- dimensional data analysis to networks and other irregular domains,” IEEE Signal Process. Mag. , vol. 30, no. 3, pp. 83–98, May 2013

work page 2013

[17] [17]

Discrete signal processing on graphs,

A. Sandryhaila and J. M. F. Moura, “Discrete signal processing on graphs,” IEEE Trans. Signal Process. , vol. 61, pp. 1644–1656, Apr. 2013

work page 2013

[18] [18]

Graph signal processing: Overview, challenges, and ap- plications,

A. Ortega, P. Frossard, J. Kova ˇcevi´c, J. M. F. Moura, and P. Van- dergheynst, “Graph signal processing: Overview, challenges, and ap- plications,” Proc. IEEE, vol. 106, no. 5, pp. 808–828, 2018

work page 2018

[19] [19]

Optimal graph-filter design and applications to distributed linear network operators,

S. Segarra, A. G. Marques, and A. Ribeiro, “Optimal graph-filter design and applications to distributed linear network operators,” IEEE Trans. Signal Process., vol. 65, pp. 4117–4131, Aug. 2017

work page 2017

[20] [20]

Convergent sequences of dense graphs i: Subgraph frequencies, metric properties and testing,

C. Borgs, J. T. Chayes, L. Lov ´asz, V . T. S ´os, and K. Vesztergombi, “Convergent sequences of dense graphs i: Subgraph frequencies, metric properties and testing,” Adv. Math. , vol. 219, no. 6, pp. 1801–1851, 2008

work page 2008

[21] [21]

Lov ´asz, Large Networks and Graph Limits , vol

L. Lov ´asz, Large Networks and Graph Limits , vol. 60, American Mathematical Society, 2012

work page 2012

[22] [22]

Transferability properties of graph neural networks,

L. Ruiz, L. F. O. Chamon, and A. Ribeiro, “Transferability properties of graph neural networks,” arXiv:2112.04629 [eess.SP]. Submitted to IEEE TSP, 2021

work page arXiv 2021

[23] [23]

How powerful are graph neural networks?,

K. Xu, W. Hu, J. Leskovec, and S. Jegelka, “How powerful are graph neural networks?,” in 7th Int. Conf. Learning Representations , New Orleans, LA, 6-9 May 2019, pp. 1–17, Assoc. Comput. Linguistics

work page 2019

[24] [24]

On the equivalence between graph isomorphism testing and function approxi- mation with gnns,

Zhengdao Chen, Soledad Villar, Lei Chen, and Joan Bruna, “On the equivalence between graph isomorphism testing and function approxi- mation with gnns,” Advances in neural information processing systems , vol. 32, 2019

work page 2019

[25] [25]

How to learn a graph from smooth signals,

Vassilis Kalofolias, “How to learn a graph from smooth signals,” in Pro- ceedings of the 19th International Conference on Artificial Intelligence and Statistics (AISTATS). PMLR, 2016, pp. 920–929

work page 2016

[26] [26]

Graph sparsification by effective resistances,

Daniel A. Spielman and Nikhil Srivastava, “Graph sparsification by effective resistances,” SIAM Journal on Computing , vol. 40, no. 6, pp. 1913–1926, 2011

work page 1913

[27] [27]

ADAM: A method for stochastic optimization,

D. P. Kingma and J. L. Ba, “ADAM: A method for stochastic optimization,” in 3rd Int. Conf. Learning Representations , San Diego, CA, 7-9 May 2015, Assoc. Comput. Linguistics. APPENDIX Proof of Prop. II.2. From the definition, we can rewrite Y = span({Skx : k = 0, . . . , K}) Since S is symmetric, there exists some E, Σ ∈ Rn×n satis- fying EET = In and Σ d...

work page 2015