GmCN: Graph Mask Convolutional Network

Beibei Wang; Bin Luo; Bo Jiang; Jin Tang

REVIEW 4 major objections 4 minor 1 cited by

GmCN: Graph Mask Convolutional Network

T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A learnable binary mask lets each node select its own aggregation neighborhood, improving GCN accuracy and graph-noise robustness.

desk verdict GmCN is a legitimate but modest GCN variant with decent experiments; the claimed theoretical optimality of the learned mask is not actually established by the derivation. read the letter →

arxiv 1910.01735 v2 pith:VGBLEI2B submitted 2019-09-04 cs.CV cs.LGcs.SIstat.ML

Bo Jiang , Beibei Wang , Jin Tang , Bin Luo This is my paper · ORCID

classification cs.CVcs.LGcs.SIstat.ML

keywords graphconvolutionalnetworksemi-supervisedlearningneighborselectionmaskmatrixstructurenoiseregularizationframeworkfeatureaggregationnodeclassification

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Graph Convolutional Networks typically aggregate every neighbor's features in every layer, on a graph structure fixed in advance. This paper argues that this full-neighborhood aggregation is not optimal for the learning task and is vulnerable to noisy edges. To fix that, it proposes GmCN, a GCN variant that learns a binary mask matrix selecting, for each node, which neighbors should participate in feature aggregation. The mask is derived from a regularized objective and learned jointly with the feature transformation by alternating updates. On five citation and co-purchase datasets the masked network outperforms GCN, GAT, GraphSAGE, and related baselines, and its accuracy degrades less when edges are randomly rewired.

What carries the argument

The load-bearing object is the mask-weighted adjacency $M\odot \hat{A}$ inside the aggregation step. The optimization objective is $$R_{\mathrm{GmCN}}(M,U)=\mathrm{Tr}[$U^{{T}}$(I-M\odot\hat{A})U]+\mu\|U-H\|$_F^{2}$+\gamma\|M\|$_F^{2}$,$$ with the binary mask relaxed to a row-stochastic nonnegative matrix. The update alternates a doubly-stochastic-like projection for $M$ (Algorithm 1) and a $T$-step truncated power iteration for $U$; the mask directly controls which graph edges carry feature information, so the network can prune unwanted connections during training.

What would settle it

On a small graph, enumerate all binary masks satisfying the row-stochastic constraint, compute $R_{\mathrm{GmCN}}$ for each, and compare the true optimum with the mask returned by Algorithm 2 after its relaxation and discretization step; if the thresholded mask's objective is substantially worse than the best binary mask, the paper's theoretical grounding for GmCN fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that a graph convolutional layer should not be tied to the given adjacency structure. GmCN instead computes $U^{(t)}=\alpha(M\odot \hat{A})U^{(t-1)}+(1-\alpha)H$, where $M$ is a per-layer learnable 0/1 mask; $M_{ij}=1$ means node $i$ aggregates features from node $j$, and $M_{ij}=0$ excludes that neighbor. The authors derive, by reinterpreting GCN's power-iteration aggregation as the solution of a regularization problem and adding a mask term, that the optimal mask is the solution of a constrained projection problem, which yields a closed-form update. They then discretize the relaxed mask and alternate mask optimization with feature aggregation, treating the whole scheme as controlled by two parameters, $\alpha$ and $\gamma$. The reported experiments claim consistent accuracy gains over fixed-graph GCN and over attention- and sampling-based alternatives, with the largest margins on the Amazon datasets and under edge perturbation.

Load-bearing premise

The theoretical interpretation assumes the continuous relaxation plus thresholding faithfully solves the original binary mask problem; the paper gives no bound or experiment measuring that discretization loss, and the threshold epsilon is not reported.

Editorial extensions

If this is right

Semi-supervised classification accuracy increases on all five datasets at 10%, 20%, and 30% labeled nodes compared with fixed-graph GCN.
The method stays competitive as the network deepens: GmCN outperforms GCN at every depth from two to seven layers on Cora.
Under random edge rewiring, GmCN's test accuracy drops less than GCN's, indicating the mask filters structure noise.
The mask update and feature aggregation come from the same regularized objective, so neighborhood selection is derived rather than set by an external heuristic.
The extra computational cost is bounded by $O(n^2)$ for the mask update and $O(n^2d)$ for aggregation in the dense worst case, with better practical efficiency when both $\hat{A}$ and $M$ are sparse.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the mask can be viewed as a learned graph-denoising operator; pruning edges during propagation is a more general mechanism than selecting neighbors for accuracy, and the same objective could be used to output a denoised graph for downstream tasks.
Editorial inference: replacing hard 0/1 discretization with a continuous mask or top-$k$ selection would make the method differentiable in the mask and could yield calibrated selection probabilities, which the current thresholding discards.
Editorial inference: because the paper reports accuracy is stable across $\gamma$ but does not report the discretization threshold $\epsilon$, the sensitivity of the final mask to thresholding remains untested; a natural experiment is to sweep $\epsilon$ and measure objective degradation and accuracy change.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Short version: this is a solid incremental paper whose experiments mostly deliver, but whose theory section overclaims. The mask idea is reasonable, and the gains over GCN, GAT, and sampling baselines are consistent on five standard datasets. But the paper does not actually prove that the learned mask solves the discrete problem it starts from.

What's new: GmCN learns a binary mask that selects which neighbors to aggregate, rather than using attention weights or fixed sampling. The mask is optimized via an alternating scheme with a projection step. That specific combination isn't in the prior work I know, and it is a sensible thing to try. The paper also does a decent job of positioning itself against GAT, GraphSAGE, CVD+PP, and the authors' own GOCN, and it includes parameter sensitivity and noise robustness experiments.

What's soft, in proportion: the central claim that the mask is optimal rests on a relaxation that isn't a relaxation. The original problem (Eq. 10-11) has binary entries with no row-sum constraint; each node may pick any subset of neighbors. The continuous problem (Eq. 12-13) forces each row to sum to one. That's a different feasible set, so the connection back to the discrete objective is broken. The threshold epsilon is never reported, and it controls whether rows become empty, single-neighbor, or multi-neighbor masks, so the final behavior is effectively determined by an invisible hyperparameter. The projection in Algorithm 1 is cited to [22], which is a doubly stochastic projection, but the stated constraint is only row-stochastic. No code is provided, so I can't tell whether the implemented projection solves the stated problem. These gaps don't kill the method—it may work fine heuristically—but they do mean the word "optimal" in the abstract and conclusion isn't supported. Also, some of the accuracy gains are within one standard deviation, though the across-dataset consistency helps.

Who it's for: researchers working on graph neural networks who want another way to do adaptive neighbor selection. The paper deserves a serious referee, but the reviewers should ask for a clarification of the relaxation, the epsilon value and sensitivity, and ideally code. It's not a breakthrough, but it's a reasonable contribution that could be made sound with revisions.

Referee Report

4 major / 4 minor

Summary. The paper proposes GmCN, a GCN variant in which a binary mask matrix M selects, for each node, a subset of neighbors to use in feature aggregation. The mask is motivated by a regularization objective (Eq. 10) that adds a sparsity penalty to a standard GCN smoothness term; the binary constraint is relaxed to a continuous row-stochastic constraint, and the resulting bi-convex problem is solved by alternating updates (Algorithm 2) followed by thresholding. Experiments on Cora, Citeseer, Cora-ML, Amazon Computers, and Amazon Photo report consistent accuracy gains over GCN, GAT, GraphSAGE, and other baselines, and a perturbation experiment shows robustness to edge noise.

Significance. The empirical results are consistently positive: GmCN obtains the best or near-best accuracy on all five datasets and label ratios, with gains of roughly 1-3 points over GCN and GAT, and the edge-perturbation experiment suggests that selective aggregation improves robustness. The regularization-based derivation is appealing in principle and differentiates the method from attention and sampling approaches. However, the paper does not provide code, and several load-bearing links between the discrete mask objective and Algorithm 2 are not established (see major comments). As it stands, the contribution is an empirically promising heuristic mask-selection method rather than a theoretically grounded optimal-mask method. If the authors repair the derivation, report the missing threshold parameter, and provide approximation analysis, the paper would be a solid contribution to the GNN literature.

major comments (4)

[§4.1, Eqs. (11)-(13)] The relaxation changes the feasible set in an unaddressed way. The original constraint M_ij ∈ {0,1} allows each row to select any subset of neighbors, including the empty set and multiple neighbors, while the relaxed constraint M1^T = 1^T, M ≥ 0 forces every row to be a probability distribution over all n entries. The subsequent thresholding with an unspecified ε can produce rows with zero, one, or many selected entries, so the final discrete M is not in general a feasible point of either problem. Moreover, under row normalization the term γ||M||^2_F no longer encourages sparsity: for a fixed row sum, the Frobenius norm is minimized by a uniform row, so the statement that 'larger γ leads to more sparse M' is not true for the relaxed problem. The authors need either a formal relationship between the discrete and relaxed problems or a change of claim to 'heuristic mask selection'.
[§4.2, Algorithms 1 and 2; §6.2] The projection step is neither derived nor matched to the stated constraint. Eq. (16) is a projection onto the row-stochastic nonnegative set, which can be solved row-wise by a simplex projection, but Algorithm 1 is described as the algorithm of [22], which is a doubly stochastic normalization, and the update in line 3 contains both left and right multiplications by 1^T1, i.e., it enforces column constraints absent from Eq. (13). The authors should either prove that the cited algorithm solves Eq. (16) or replace it with the correct projection; as written, the mask returned by Algorithm 2 is not shown to minimize Eq. (16) even before discretization.
[§4.2, Eqs. (17)-(18); Algorithm 2] The alternating scheme is only an approximation, and the label 'optimal' is unsupported. Step 2 replaces the exact minimizer of Eq. (17) by T = 3 power iterations, and the outer loop is truncated at K = 4, L = 3, T = 3; no error bound, convergence criterion, or ablation on K, L, T is provided. Since M and U are never optimized to convergence, the output is not shown to be a stationary point of Eq. (12). The paper should either provide approximation guarantees or describe GmCN as an approximate/heuristic algorithm and temper the word 'optimal' in the abstract, Section 5, and conclusion.
[§6.2; Algorithm 2] The discretization threshold ε is a free parameter of the method but is never reported or analyzed. The text lists settings for α, γ, K, L, T but not ε; the parameter analysis in Tables 4-5 covers only α and γ, and Figure 1 covers only depth and noise level. Since ε determines whether a row becomes empty, single-neighbor, or multi-neighbor after thresholding, the reported experiments are not reproducible without this value, and there is no evidence that thresholding the continuous solution preserves closeness to an optimum of Eq. (10). Please report ε, add a sensitivity study, and quantify the discretization loss, e.g., by comparing the objective value before and after thresholding.

minor comments (4)

[§4.1] There are two typographical errors: 'Forbnuous norm' should be 'Frobenius norm', and 'combinational' should be 'combinatorial'.
[§6.1] The Cora dataset is described as having six classes, but the standard Cora dataset has seven classes; please correct this and verify the corresponding statistics.
[§6.5] The perturbation study compares GmCN only against GCN; including one or two additional baselines, such as GAT or APPNP, would make the robustness claim more convincing.
[Tables 1-3] Several differences between GmCN and the runner-up are within one standard deviation; reporting paired significance tests or confidence intervals over the five splits would strengthen the claim that GmCN is consistently better.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the mask update is the minimizer of an explicitly stated variational objective, not a fitted parameter renamed as a prediction.

full rationale

The GmCN derivation chain is self-contained in the required sense. Section 4.1 invokes the known regularization interpretation of GCN power iteration, Eq. (9), with the proof attributed to the independent prior work [23] (Zhou et al.); refs. [7,8] are additional pointers and are not load-bearing. Eq. (10) is then a genuine variational extension obtained by replacing \hat A with M⊙\hat A and adding a penalty on M, not an objective defined in terms of the algorithm's output. The M-subproblem is algebraically reduced to the projection problem Eq. (16), and Algorithm 2 implements that projection together with the U-subproblem power iteration of Eq. (18). Thus the learned mask is the minimizer of a stated objective rather than a fitted quantity renamed as a prediction. The concerns raised about the paper—that the relaxation in Eq. (13) adds a row-sum constraint absent from the discrete problem, that the discretization threshold \epsilon is unreported, and that the projection is cited to a doubly stochastic algorithm while the constraint is only row-stochastic—are correctness/validation gaps, not circular reductions. No step in the paper reduces to its own inputs by construction, and no central claim depends on a self-citation chain.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the validity of the relaxed-and-thresholded mask optimization, on the borrowed regularization framework from prior work (including self-citations), and on a set of hyperparameters alpha, gamma, and epsilon, the last of which is not reported. The projection algorithm is imported from the literature without proof. These are the main ingredients the reader must accept to trust the theoretical interpretation.

free parameters (4)

alpha = 0.8
Controls the balance between neighbor aggregation and self-feature retention in the propagation update; tuned on validation to obtain best average performance.
gamma = 0.001
Sparsity penalty for the mask; chosen by hand to obtain best average performance.
epsilon = not specified
Threshold for discretizing the relaxed mask into binary values; no value reported in experiments, leaving behavior unspecified.
K, L, T = 4, 3, 3
Maximum inner iterations for mask and aggregation updates; chosen for efficiency, affects approximation quality.

assumptions (4)

standard math The GCN feature aggregation in Eq. (3) is a T-step power iteration solution to the regularization framework in Eq. (9) (Zhou et al. 2004; Jiang et al. 2019).
Borrowed from cited prior work, including two self-citations; it is the starting point for the GmCN objective.
ad hoc to paper The binary mask constraint M_ij in {0,1} can be relaxed to a continuous row-stochastic constraint M1^T = 1^T, M >= 0, and the resulting solution can be post-discretized without significant loss.
This relaxation is introduced in Section 4.1 and used in Algorithm 1, but no analysis shows the discretized mask remains near-optimal.
standard math The projection algorithm from Zass and Shashua (2007) solves the constrained projection problem in Eq. (16).
The paper cites this algorithm without detailing or proving it; it is a standard method for doubly stochastic projection.
ad hoc to paper The amortized aggregation step in Eq. (18) with T=3 is a sufficient approximation of the U-update in Eq. (17).
The paper uses T=3 truncated iterations and K=4 outer loops, assuming these suffice for convergence, without empirical support.

how reviews work

0 comments

Cite this review

Pith. "Pith review of GmCN: Graph Mask Convolutional Network." pith.science (2026). https://pith.science/paper/VGBLEI2B

@misc{pith2026191001735,
  author       = {Pith},
  title        = {Pith review of: GmCN: Graph Mask Convolutional Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VGBLEI2B}},
  note         = {Machine review of arXiv:1910.01735}
}

read the original abstract

Graph Convolutional Networks (GCNs) have shown very powerful for graph data representation and learning tasks. Existing GCNs usually conduct feature aggregation on a fixed neighborhood graph in which each node computes its representation by aggregating the feature representations of all its neighbors which is biased by its own representation. However, this fixed aggregation strategy is not guaranteed to be optimal for GCN based graph learning and also can be affected by some graph structure noises, such as incorrect or undesired edge connections. To address these issues, we propose a novel Graph mask Convolutional Network (GmCN) in which nodes can adaptively select the optimal neighbors in their feature aggregation to better serve GCN learning. GmCN can be theoretically interpreted by a regularization framework, based on which we derive a simple update algorithm to determine the optimal mask adaptively in GmCN training process. Experiments on several datasets validate the effectiveness of GmCN.

Figures

Figures reproduced from arXiv: 1910.01735 by the authors.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Semi-supervised Learning with Adaptive Neighborhood Graph Propagation Network
cs.CV 2019-08 conditional novelty 4.0 of 10

ANGPN jointly learns an adaptive neighborhood graph and propagates node features through it, improving semi-supervised classification accuracy by a few points over fixed-graph GCN on SVHN, CIFAR, 20News, and CoraML.

Reference graph

Works this paper leans on

24 extracted references · 17 canonical work pages · cited by 1 Pith paper

[22]

Zass and A

R. Zass and A. Shashua. Doubly stochastic normalization for spectral clustering. In Advances in neural information processing systems, pages 1569–1576, 2007

work page 2007
[1]

Bruna, W

J. Bruna, W. Zaremba, A. Szlam, and Y . LeCun. Spectral networks and locally connected networks on graphs. In International Conference on Learning Representations, 2014

work page 2014
[2]

J. Chen, J. Zhu, and L. Song. Stochastic training of graph convolutional networks with variance reduction. In International Conference on Machine Learning, pages 942–950, 2018

work page 2018
[3]

Defferrard, X

M. Defferrard, X. Bresson, and P. Vandergheynst. Convolutional neural networks on graphs with fast localized spectral ﬁltering. In Advances in Neural Information Processing Systems, pages 3844–3852, 2016

work page 2016
[4]

Glorot and Y

X. Glorot and Y . Bengio. Understanding the difﬁculty of training deep feedforward neural networks. In International conference on artiﬁcial intelligence and statistics, pages 249–256, 2010

work page 2010
[5]

Hamilton, Z

W. Hamilton, Z. Ying, and J. Leskovec. Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems, pages 1024–1034, 2017. 8

work page 2017
[6]

Henaff, J

M. Henaff, J. Bruna, and Y . LeCun. Deep convolutional networks on graph-structured data. arXiv preprint arXiv:1506.05163, 2015

arXiv 2015
[7]

Jiang, D

B. Jiang, D. Lin, J. Tang, and B. Luo. Data representation and learning with graph diffusion- embedding networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10414–10423, 2019

work page 2019

Show all 24 references

[8]

Jiang, Z

B. Jiang, Z. Zhang, J. Tang, and B. Luo. Graph optimized convolutional networks. arXiv:1904.11883, 2019

1904 arXiv
[9]

D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015

2015
[10]

T. N. Kipf and M. Welling. Semi-supervised classiﬁcation with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016

2016 arXiv
[11]

Klicpera, A

J. Klicpera, A. Bojchevski, and S. Günnemann. Predict then propagate: Graph neural networks meet personalized pagerank. 2019

2019
[12]

Y . Li, R. Yu, C. Shahabi, and Y . Liu. Diffusion convolutional recurrent neural network: Data- driven trafﬁc forecasting. In International Conference on Learning Representations (ICLR ’18), 2018

2018
[13]

McAuley, C

J. McAuley, C. Targett, Q. Shi, and A. Van Den Hengel. Image-based recommendations on styles and substitutes. In Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 43–52, 2015

2015
[14]

A. K. McCallum, K. Nigam, J. Rennie, and K. Seymore. Automating the construction of internet portals with machine learning. Information Retrieval, 3(2):127–163, 2000

2000
[15]

Monti, D

F. Monti, D. Boscaini, J. Masci, E. Rodola, J. Svoboda, and M. M. Bronstein. Geometric deep learning on graphs and manifolds using mixture model cnns. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5423–5434, 2017

2017
[16]

Perozzi, R

B. Perozzi, R. Al-Rfou, and S. Skiena. Deepwalk: Online learning of social representations. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 701–710, 2014

2014
[17]

Ruoyu, W

L. Ruoyu, W. Sheng, Z. Feiyun, and H. Junzhou. Adaptive graph convolutional neural networks. In AAAI Conference on Artiﬁcial Intelligence, pages 3546–3553, 2018

2018
[18]

P. Sen, G. Namata, M. Bilgic, L. Getoor, B. Galligher, and T. Eliassi-Rad. Collective classiﬁca- tion in network data. AI magazine, 29(3):93, 2008

2008
[19]

Shchur, M

O. Shchur, M. Mumme, A. Bojchevski, and S. Günnemann. Pitfalls of graph neural network evaluation. arXiv preprint arXiv:1811.05868, 2018

2018 arXiv
[20]

Velickovic, G

P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y . Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017

2017 arXiv
[21]

Veliˇckovi´c, W

P. Veliˇckovi´c, W. Fedus, W. L. Hamilton, P. Liò, Y . Bengio, and R. D. Hjelm. Deep graph infomax. In ICLR, 2019

2019
[23]

D. Zhou, O. Bousquet, T. N. Lal, J. Weston, and B. Schölkopf. Learning with local and global consistency. In Advances in neural information processing systems, pages 321–328, 2004

2004
[24]

J. Zhou, G. Cui, Z. Zhang, C. Yang, Z. Liu, and M. Sun. Graph neural networks: A review of methods and applications. arXiv preprint arXiv:1812.08434, 2018. 9

2018 arXiv

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [22]

Zass and A

R. Zass and A. Shashua. Doubly stochastic normalization for spectral clustering. In Advances in neural information processing systems, pages 1569–1576, 2007

work page 2007

[2] [1]

Bruna, W

J. Bruna, W. Zaremba, A. Szlam, and Y . LeCun. Spectral networks and locally connected networks on graphs. In International Conference on Learning Representations, 2014

work page 2014

[3] [2]

J. Chen, J. Zhu, and L. Song. Stochastic training of graph convolutional networks with variance reduction. In International Conference on Machine Learning, pages 942–950, 2018

work page 2018

[4] [3]

Defferrard, X

M. Defferrard, X. Bresson, and P. Vandergheynst. Convolutional neural networks on graphs with fast localized spectral ﬁltering. In Advances in Neural Information Processing Systems, pages 3844–3852, 2016

work page 2016

[5] [4]

Glorot and Y

X. Glorot and Y . Bengio. Understanding the difﬁculty of training deep feedforward neural networks. In International conference on artiﬁcial intelligence and statistics, pages 249–256, 2010

work page 2010

[6] [5]

Hamilton, Z

W. Hamilton, Z. Ying, and J. Leskovec. Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems, pages 1024–1034, 2017. 8

work page 2017

[7] [6]

Henaff, J

M. Henaff, J. Bruna, and Y . LeCun. Deep convolutional networks on graph-structured data. arXiv preprint arXiv:1506.05163, 2015

arXiv 2015

[8] [7]

Jiang, D

B. Jiang, D. Lin, J. Tang, and B. Luo. Data representation and learning with graph diffusion- embedding networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10414–10423, 2019

work page 2019