REVIEW 3 major objections 5 minor 41 references
Feature Construction Using Network Control Theory and Rank Encoding for Graph Machine Learning
T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read Features built from average controllability and rank encoding improve GNN graph classification on unattributed social networks.
desk verdict A competent, incremental feature-construction paper whose central claim is broader than the evidence: the controllability computation is underspecified, gains are dataset-dependent, and the new rank encoding is standard histogram binning. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is average controllability, the diagonal of the controllability Gramian $W(A,B)$ of the dynamics $\dot{x}(t)=A x(t)+B u(t)$; the paper computes a finite-horizon version by numerical integration over $[0, T=1]$ with step $0.001$, which avoids the stability requirement of the infinite-horizon Lyapunov equation. The second mechanism is rank encoding: build a $k$-bin histogram of a scalar metric, assign each node to its bin, and represent the node by a one-hot vector on that bin index. The NCT-EFA variant concatenates raw metric values for average controllability, closeness, betweenness, and eigenvector centrality into the initial feature matrix the GNN consumes.
What would settle it
Recompute the finite-horizon Gramian on GitHub Stargazers with different horizons (e.g., $T=0.1$ and $T=10$) and step sizes (e.g., $0.01$ and $0.0001$), then check whether node rankings stay nearly the same and whether GraphSAGE's ROC AUC advantage (73.9% vs 68.7%) survives; a material change in either would show the gains depend on the integration settings rather than on the controllability concept.
Extended reading notes
Core claim
On its own terms, the paper's central claim is that average controllability — the diagonal of the controllability Gramian for the linear dynamics $\dot{x}(t) = A x(t) + B u(t)$ with actuation at every node — is a usable and informative node feature for unattributed graphs. Combined with closeness, betweenness, and eigenvector centrality into a composite vector (NCT-EFA), it improves graph-classification ROC AUC over one-hot degree encoding in most settings tested. The paper further claims that rank encoding, which maps any scalar node metric into a fixed-dimensional one-hot vector through a $k$-bin histogram, outperforms one-hot degree encoding, and that concatenating rank-encoded versions of several metrics gives the most consistent gains.
Load-bearing premise
The load-bearing premise is that the finite-horizon controllability Gramian computed over $[0, T=1]$ with step $0.001$ produces a stable and meaningful ranking of node influence, and the paper gives no sensitivity analysis for those settings.
Editorial extensions
If this is right
- On unattributed social graphs whose structure carries the label, controllability- and centrality-based initial features should replace one-hot degree encoding as the default feature initialization for graph classification.
- Because rank encoding fixes the feature dimension by the bin count $k$ rather than by the maximum degree, it should scale better on power-law networks with huge degree ranges.
- The reported bin-size experiments imply that $k=30$ is a reasonable default on Reddit Threads and GitHub Stargazers, while bin count matters little on Twitch Egos and Deezer Egos.
- The Deezer Egos results imply the method's gains are bounded by how much label information the graph's topology actually contains; on weak-signal tasks even the best encoding stays near chance.
Reading between the lines
- My inference: rank encoding is a generic transform; the same $k$-bin one-hot construction could be applied to any scalar node score, and a natural test is link prediction or node classification, where the merits of a fixed-dimensional feature are easier to isolate than in graph classification.
- My inference: the headline gain may come as much from histogram smoothing of skewed metric distributions as from controllability itself; the paper does not run the control experiment of rank-encoding degree or random scores with the same binning, which would separate encoding from metric.
- My inference: since the controllability Gramian and the graph spectrum are mathematically related, much of what average controllability provides may be reproducible from cheaper spectral features; if so, the durable contribution is the encoding, not the control-theoretic computation.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes two feature-construction strategies for GNN-based graph classification on unattributed social networks: (i) NCT-EFA, which concatenates average controllability with closeness, betweenness, and eigenvector centrality as node features, and (ii) rank encoding, which maps any scalar graph metric to a fixed-dimensional one-hot vector via histogram binning. The methods are evaluated on Reddit Threads, GitHub Stargazers, Twitch Egos, and Deezer Egos with six GNN architectures, using one-hot degree encoding as baseline. The main reported result is that rank-encoded average controllability improves GraphSAGE ROC AUC from 68.7% to 73.9% on GitHub Stargazers, with smaller gains on Reddit and Twitch and mixed results on Deezer Egos.
Significance. If the claims are robust, the paper offers a simple, type-agnostic encoding scheme that converts scalar structural metrics into fixed-dimensional GNN features, which is useful for unattributed graphs where node attributes are missing. The breadth of the evaluation—six GNNs and four datasets—and the explicit discussion of computational complexity are strengths. The core numerical claims, however, rest on an underspecified and internally inconsistent finite-horizon controllability Gramian computation, and the reported gains are not supported by significance tests and are contradicted on some datasets. The conceptual contribution is therefore interesting and potentially publishable, but the evidence presented in the current manuscript is not yet sufficient to support the abstract's claims of significant and consistent improvement.
major comments (3)
- [3.2.1, Eqs. (1)-(3)] The mathematical definition of average controllability is internally inconsistent. Eq. (2) defines the Gramian with e^{-Aτ} and Eq. (3) states the Lyapunov equation AW + WA^T + BB^T = 0, which correspond to stable dynamics with negative feedback, not to Eq. (1)'s ẋ = Ax + Bu. The finite-horizon computation is described only verbally as 'simulating the system dynamics over discrete time steps and integrating the resulting matrix sequence,' with T=1 and step=0.001; no discretization formula is given. For an undirected adjacency matrix with positive spectral radius, the correct finite-horizon Gramian ∫_0^T e^{Aτ}BB^T e^{A^Tτ}dτ grows rapidly with τ, so the diagonal entries are highly sensitive to the choice of T and the integration rule. Since every proposed feature depends on this quantity, and rank encoding turns any change in node ordering into different one-hot features, the reported GraphSAGE gain on GitHub Stargazers could be an artifact of one arbitrary integration setting. Please provide the exact discretization, or use a well-defined stable formulation, and include a sensitivity analysis over T and step size.
- [Abstract, 4.3, 4.4] The claim that incorporating average controllability 'significantly improves' GNN performance is not supported by the reported tables. On Deezer Egos, degree encoding outperforms NCT-EFA in most model/dataset cells (e.g., Table 2: GraphSAGE deg 53.29 vs NCT-EFA 51.68; GAT deg 53.30 vs NCT-EFA 52.70), and in Table 3 AC encoding underperforms degree for GraphSAGE and UniMP. On Twitch Egos, the gains are within one standard deviation (e.g., Table 2: GraphConv 65.26 vs 65.70). The paper reports no significance tests or paired comparisons across cross-validation folds. Please add appropriate statistical testing and revise the abstract and conclusion to state where the improvements are consistent and where they are not.
- [4.4 and Figures 3-4] The claim that k=30 achieves the best results in 5 out of 6 cases on Reddit and 4 out of 6 on GitHub is not verifiable from the paper: the figures do not report error bars or exact numerical values, and the y-axis ranges are truncated, making differences appear larger than they may be. Moreover, the effect of k is not statistically tested, so statements such as 'k has no substantial impact' (Deezer and Twitch) are informal. Please report exact values and error bars, or move this analysis to supplementary material with quantitative comparisons.
minor comments (5)
- [Table 2] The table header is garbled (e.g., 'GATdeg NCT-EFAdeg NCT-EFA...'), which makes it difficult to read which column corresponds to which model and feature set. Please reformat the table with clear per-model deg and NCT-EFA columns.
- [Section 3.2.2] The binning example is inconsistent: Bin 1 is defined as [0.0, 0.3), but Node 2 with value 0.3 is said to fall into Bin 1. Specify half-open intervals consistently, e.g., [0.0, 0.3), [0.3, 0.6), [0.6, 1.0], or adjust the example.
- [Section 3.2.1] One-hot degree encoding is described as an n-dimensional vector, but the standard construction uses a dimension equal to the maximum degree plus one, not the number of nodes. Please clarify the exact implementation, since this affects the comparison with the fixed-dimensional rank encoding.
- [6] The heading 'Acknolwdgements' contains a typo; it should be 'Acknowledgements'.
- [General] No code or data release is mentioned. Given that the central feature computation is underspecified, a reproducibility statement or code link would materially strengthen the paper.
Circularity Check
No significant circularity: node features are computed from graph topology before any training, and all headline comparisons are against external benchmark datasets.
full rationale
The paper's derivation chain is self-contained with respect to circularity. The proposed features (average controllability, closeness, betweenness, eigenvector centrality, and the rank-encoded histograms) are computed from the graph adjacency matrix before any classifier is trained; they are not fitted to the classification labels, and no equation in the paper defines a feature in terms of the prediction target or vice versa. The rank encoding method is described explicitly as histogram binning followed by one-hot assignment, and its parameters (number of bins, histogram range) are fixed a priori from the metric values, not learned from the labels. Comparisons against one-hot degree encoding and other baselines use external public datasets (Reddit Threads, GitHub Stargazers, Twitch Egos, Deezer Egos) and standard GNN architectures, so the reported AUC improvements are empirical outcomes rather than consequences of a fitted parameter being renamed as a prediction. The paper cites prior work by the same authors (e.g., [12], [13]) for context and as a predecessor of the approach, but the central evaluation does not rest on an unverified uniqueness theorem or on a self-citation chain that forbids alternatives. The finite-horizon controllability Gramian is described with a sign inconsistency and without a discretization formula, and the lack of sensitivity analysis for T and step is a genuine reproducibility and robustness concern; however, that concern is about correctness of the numerical computation, not about circularity, because the Gramian values are still computed from the graph structure rather than derived from the classification labels. Consequently, there is no circular step that reduces any claimed result to its own input by construction.
Assumptions & free parameters
free parameters (4)
- Finite-horizon integration horizon T =
1
- Integration step size =
0.001
- Number of histogram bins k =
10 (main tables); 20 and 30 in sweep
- GNN training hyperparameters =
3 layers, 64 hidden units, lr=1e-4, weight decay=5e-2, 100 epochs
assumptions (4)
- ad hoc to paper Average controllability is computed from the diagonal of the finite-horizon controllability Gramian with T=1 and step=0.001, and this quantity is a meaningful and stable node feature.
- domain assumption The linear diffusion model x-dot = Ax + Bu with full control B=I is an appropriate model of social-network influence.
- domain assumption Structural topology alone, without node attributes, carries enough signal for the four classification tasks.
- standard math Standard centrality definitions for closeness, betweenness, and eigenvector centrality are computed as described.
Cite this review
Pith. "Pith review of Feature Construction Using Network Control Theory and Rank Encoding for Graph Machine Learning." pith.science (2026). https://pith.science/paper/3ESZFPJW
@misc{pith2026250715195,
author = {Pith},
title = {Pith review of: Feature Construction Using Network Control Theory and Rank Encoding for Graph Machine Learning},
year = {2026},
howpublished = {\url{https://pith.science/paper/3ESZFPJW}},
note = {Machine review of arXiv:2507.15195}
}
read the original abstract
In this article, we utilize the concept of average controllability in graphs, along with a novel rank encoding method, to enhance the performance of Graph Neural Networks (GNNs) in social network classification tasks. GNNs have proven highly effective in various network-based learning applications and require some form of node features to function. However, their performance is heavily influenced by the expressiveness of these features. In social networks, node features are often unavailable due to privacy constraints or the absence of inherent attributes, making it challenging for GNNs to achieve optimal performance. To address this limitation, we propose two strategies for constructing expressive node features. First, we introduce average controllability along with other centrality metrics (denoted as NCT-EFA) as node-level metrics that capture critical aspects of network topology. Building on this, we develop a rank encoding method that transforms average controllability or any other graph-theoretic metric into a fixed-dimensional feature space, thereby improving feature representation. We conduct extensive numerical evaluations using six benchmark GNN models across four social network datasets to compare different node feature construction methods. Our results demonstrate that incorporating average controllability into the feature space significantly improves GNN performance. Moreover, the proposed rank encoding method outperforms traditional one-hot degree encoding, improving the ROC AUC from 68.7% to 73.9% using GraphSAGE on the GitHub Stargazers dataset, underscoring its effectiveness in generating expressive and efficient node representations.
Figures
Reference graph
Works this paper leans on
-
[1]
Controllability of structural brain networks,
S. Gu, F. Pasqualetti, M. Cieslak, Q. K. Telesford, A. B. Yu, A. E. Kahn, J. D. Medaglia, J. M. Vettel, M. B. Miller, S. T. Grafton et al., “Controllability of structural brain networks,” Nature communications, vol. 6, no. 1, p. 8414, 2015
work page 2015
-
[2]
Friedland, Control system design: an introduction to state-space methods
B. Friedland, Control system design: an introduction to state-space methods. Courier Corporation, 2012
work page 2012
-
[3]
Controllability metrics, limitations and algorithms for complex net- works,
F. Pasqualetti, S. Zampieri, and F. Bullo, “Controllability metrics, limitations and algorithms for complex net- works,” IEEE Transactions on Control of Network Systems, vol. 1, no. 1, pp. 40–52, 2014
work page 2014
-
[4]
W. L. Hamilton, Graph representation learning. Morgan & Claypool Publishers, 2020
work page 2020
-
[5]
Semi-supervised classification with graph convolutional networks,
T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” inInternational Conference on Learning Representations, 2017
2017
-
[6]
Inductive representation learning on large graphs,
W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” Advances in neural information processing systems, vol. 30, 2017
2017
-
[7]
How powerful are graph neural networks?
K. Xu, W. Hu, J. Leskovec, and S. Jegelka, “How powerful are graph neural networks?” in International Con- ference on Learning Representations, 2019
work page 2019
-
[8]
Contributions to the theory of optimal control,
R. E. Kalman et al., “Contributions to the theory of optimal control,” Bol. soc. mat. mexicana, vol. 5, no. 2, pp. 102–119, 1960
work page 1960
Show all 41 references
-
[9]
Diffusion improves graph learning,
J. Gasteiger, S. Weißenberger, and S. Günnemann, “Diffusion improves graph learning,” Advances in neural information processing systems, vol. 32, 2019
2019
-
[10]
Centrality in social networks: Conceptual clarification,
L. C. Freeman et al., “Centrality in social networks: Conceptual clarification,” Social network: critical concepts in sociology. Londres: Routledge, vol. 1, pp. 238–263, 2002
2002
-
[11]
Network controllability in transmodal cortex predicts positive psychosis spectrum symptoms,
L. Parkes, T. M. Moore, M. E. Calkins, M. Cieslak, D. R. Roalf, D. H. Wolf, R. C. Gur, R. E. Gur, T. D. Sat- terthwaite, and D. S. Bassett, “Network controllability in transmodal cortex predicts positive psychosis spectrum symptoms,” Biological Psychiatry, vol. 90, no. 6, pp. ...
2021
-
[12]
Improving graph machine learning perfor- mance through feature augmentation based on network control theory,
A. Said, O. U. Ahmad, W. Abbas, M. Shabbir, and X. Koutsoukos, “Improving graph machine learning perfor- mance through feature augmentation based on network control theory,” in2024 32nd Mediterranean Conference on Control and Automation (MED). IEEE, 2024, pp. 322–327
2024
-
[13]
Network controllability perspectives on graph representation,
——, “Network controllability perspectives on graph representation,” IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 8, pp. 4116–4128, 2023
2023
-
[14]
Using network control theory to study the dynamics of the structural connectome,
L. Parkes, J. Z. Kim, J. Stiso, J. K. Brynildsen, M. Cieslak, S. Covitz, R. E. Gur, R. C. Gur, F. Pasqualetti, R. T. Shinohara et al., “Using network control theory to study the dynamics of the structural connectome,” bioRxiv, 2023
2023
-
[15]
Linear dynamics and control of brain networks,
J. Z. Kim and D. S. Bassett, “Linear dynamics and control of brain networks,”Neural Engineering, pp. 497–518, 2020
2020
-
[16]
Control the gnn: Utilizing neural controller with lyapunov stability for test-time feature reconstruction,
J. Yang, R. Ding, F. Ji, H. Wang, and L. Xie, “Control the gnn: Utilizing neural controller with lyapunov stability for test-time feature reconstruction,” arXiv preprint arXiv:2410.09708, 2024
2024 arXiv
-
[17]
A survey on graph kernels,
N. M. Kriege, F. D. Johansson, and C. Morris, “A survey on graph kernels,” Applied Network Science , vol. 5, no. 1, pp. 1–42, 2020
2020
-
[18]
Wasserstein weisfeiler-lehman graph kernels,
M. Togninalli, E. Ghisu, F. Llinares-López, B. Rieck, and K. Borgwardt, “Wasserstein weisfeiler-lehman graph kernels,” Advances in neural information processing systems, vol. 32, 2019
2019
-
[19]
The multiscale laplacian graph kernel,
R. Kondor and H. Pan, “The multiscale laplacian graph kernel,” Advances in neural information processing systems, vol. 29, 2016. 12 A PREPRINT - SEPTEMBER 3, 2025
2016
-
[20]
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,” Advances in neural information processing systems, vol. 29, 2016
2016
-
[21]
Spectral networks and locally connected networks on graphs,
J. Bruna, W. Zaremba, A. Szlam, and Y . Lecun, “Spectral networks and locally connected networks on graphs,” in International Conference on Learning Representations (ICLR2014), 2014, pp. http–openreview
2014
-
[22]
Residual gated graph convnets,
X. Bresson and T. Laurent, “Residual gated graph convnets,” arXiv preprint arXiv:1711.07553, 2017
2017 arXiv
-
[23]
Graph attention networks,
P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Liò, and Y . Bengio, “Graph attention networks,” in International Conference on Learning Representations, 2018
2018
-
[24]
Weisfeiler and leman go neural: Higher-order graph neural networks,
C. Morris, M. Ritzert, M. Fey, W. L. Hamilton, J. E. Lenssen, G. Rattan, and M. Grohe, “Weisfeiler and leman go neural: Higher-order graph neural networks,” inProceedings of the AAAI conference on artificial intelligence, vol. 33, no. 01, 2019, pp. 4602–4609
2019
-
[25]
Graph transformer networks,
S. Yun, M. Jeong, R. Kim, J. Kang, and H. J. Kim, “Graph transformer networks,”Advances in neural information processing systems, vol. 32, 2019
2019
-
[26]
Hierarchical graph representation learning with differentiable pooling,
Z. Ying, J. You, C. Morris, X. Ren, W. Hamilton, and J. Leskovec, “Hierarchical graph representation learning with differentiable pooling,” Advances in neural information processing systems, vol. 31, 2018
2018
-
[27]
Edge contraction pooling for graph neural networks,
F. Diehl, “Edge contraction pooling for graph neural networks,” CoRR, 2019
2019
-
[28]
A comprehensive survey on graph neural networks,
Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and S. Y . Philip, “A comprehensive survey on graph neural networks,” IEEE transactions on neural networks and learning systems, no. 1, 2020
2020
-
[29]
A simple yet effective baseline for non-attributed graph classification,
C. Cai and Y . Wang, “A simple yet effective baseline for non-attributed graph classification,” arXiv preprint arXiv:1811.03508, 2018
2018 arXiv
-
[30]
Attention models in graphs: A survey,
J. B. Lee, R. A. Rossi, S. Kim, N. K. Ahmed, and E. Koh, “Attention models in graphs: A survey,” ACM Transactions on Knowledge Discovery from Data (TKDD), vol. 13, no. 6, pp. 1–25, 2019
2019
-
[31]
Laplacian eigenmaps for dimensionality reduction and data representation,
M. Belkin and P. Niyogi, “Laplacian eigenmaps for dimensionality reduction and data representation,” Neural computation, vol. 15, no. 6, pp. 1373–1396, 2003
2003
-
[32]
Benchmarking graph neural networks,
V . P. Dwivedi, C. K. Joshi, A. T. Luu, T. Laurent, Y . Bengio, and X. Bresson, “Benchmarking graph neural networks,” Journal of Machine Learning Research, vol. 24, no. 43, pp. 1–48, 2023
2023
-
[33]
struc2vec: Learning node representations from structural identity,
L. F. Ribeiro, P. H. Saverese, and D. R. Figueiredo, “struc2vec: Learning node representations from structural identity,” in Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, 2017, pp. 385–394
2017
-
[34]
role2vec: Role-based network embeddings,
N. K. Ahmed, R. A. Rossi, J. B. Lee, T. L. Willke, R. Zhou, X. Kong, and H. Eldardiry, “role2vec: Role-based network embeddings,” Proc. DLG KDD, pp. 1–7, 2019
2019
-
[35]
Learning structural node embeddings via diffusion wavelets,
C. Donnat, M. Zitnik, D. Hallac, and J. Leskovec, “Learning structural node embeddings via diffusion wavelets,” in Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, 2018, pp. 1320–1329
2018
-
[36]
A unified view on graph neural networks as graph signal denoising,
Y . Ma, X. Liu, T. Zhao, Y . Liu, J. Tang, and N. Shah, “A unified view on graph neural networks as graph signal denoising,” in Proceedings of the 30th ACM international conference on information & knowledge management, 2021, pp. 1202–1211
2021
-
[37]
Control-based graph embeddings with data augmentation for contrastive learning,
O. U. Ahmad, A. Said, M. Shabbir, X. Koutsoukos, and W. Abbas, “Control-based graph embeddings with data augmentation for contrastive learning,” in 2024 American Control Conference (ACC). IEEE, 2024, pp. 27–32
2024
-
[38]
Social network analysis: Methods and applications,
S. Wasserman, “Social network analysis: Methods and applications,” The Press Syndicate of the University of Cambridge, 1994
1994
-
[39]
Karate Club: An API Oriented Open-source Python Framework for Unsupervised Learning on Graphs,
B. Rozemberczki, O. Kiss, and R. Sarkar, “Karate Club: An API Oriented Open-source Python Framework for Unsupervised Learning on Graphs,” in Proceedings of the 29th ACM International Conference on Information and Knowledge Management (CIKM ’20). ACM, 2020, p. 3125–3132
2020
-
[40]
Masked label prediction: Unified message passing model for semi-supervised classification,
Y . Shi, Z. Huang, S. Feng, H. Zhong, W. Wang, and Y . Sun, “Masked label prediction: Unified message passing model for semi-supervised classification,” in Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence. International Joint Conferences o...
2021
-
[41]
An end-to-end deep learning architecture for graph classification,
M. Zhang, Z. Cui, M. Neumann, and Y . Chen, “An end-to-end deep learning architecture for graph classification,” in Proceedings of the AAAI conference on artificial intelligence, vol. 32, no. 1, 2018. 13
2018
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.