AdaGCN: Adaboosting Graph Convolutional Networks into Deep Models

Ke Sun; Zhanxing Zhu; Zhouchen Lin

REVIEW 3 major objections 4 minor 21 references

AdaGCN: Adaboosting Graph Convolutional Networks into Deep Models

T0 review · 3 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 AdaGCN recasts depth in graph networks as an AdaBoost ensemble over neighbor hops and reports state-of-the-art node classification at near-constant per-epoch cost.

desk verdict A new AdaBoost-over-hops GNN with careful experiments and two overstated claims: SOTA consistency and per-epoch cost advantage. read the letter →

arxiv 1908.05081 v3 pith:6MO6NX7S submitted 2019-08-14 cs.LG stat.ML

Ke Sun , Zhanxing Zhu , Zhouchen Lin This is my paper · ORCID

classification cs.LGstat.ML

keywords AdaGCNgraphconvolutionalnetworksBoostSAMME.RnodeclassificationoversmoothingpersonalizedPageRankmulti-hopneighboraggregation

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

AdaGCN proposes that the right way to make a graph neural network deep is not to stack convolutional layers but to run one compact classifier many times over increasingly distant neighbor aggregates and boost the results. At round $l$, the classifier sees the precomputed $l$-hop aggregate $\hat{A}^l X$; a weighted error rate then sets an AdaBoost combination weight and reweights the training nodes so later rounds focus on nodes earlier hops misclassified. The paper reports that accuracy keeps improving as layers are added, that the method beats standard GCN and personalized-PageRank propagation on citation and social graphs across label rates, and that the gains are largest when labels are scarce. It also reports a computational dividend: the sparse adjacency matrix enters only through $L$ precomputed products, so per-epoch training cost is nearly flat as depth grows. The paper's two propositions frame AdaGCN as an adaptive generalization of APPNP and as a model that can represent arbitrary layer-wise neighborhood mixing.

What carries the argument

The load-bearing object is the sequence of graph-power feature matrices $\hat{X}^{(l)} = \hat{A}^l X$, where $\hat{A}$ is the normalized adjacency matrix with self-loops; these matrices are computed once and each one is handed to the round-$l$ classifier. The mechanism around them is the SAMME.R loop: fit $f_\theta^{(l)}$ to $\hat{X}^{(l)}$ under the current node weights, form the weighted probability estimates $p_k^{(l)}$, set the combination scores $h_k^{(l)}$ as above, then reweight node $i$ by $w_i \leftarrow w_i \exp\left(-\frac{K-1}{K} y_i^\top \log p^{(l)}(x_i)\right)$ and renormalize. The same two-layer architecture is reused each round with the previous round's parameters as initialization, which makes the ensemble RNN-like while keeping the sparse adjacency matrix out of the network's forward and backward passes.

What would settle it

Run AdaGCN on Cora-ML twice, once with adaptive SAMME.R weights and once with the same per-layer weights frozen to APPNP's exponential-decay schedule while keeping the shared two-layer classifier; if the frozen version matches or beats the adaptive version, the reported accuracy gains are not coming from AdaBoost adaptivity.

Watch

Extended reading notes

Core claim

The paper's central claim is that depth in graph learning can be replaced by an AdaBoost-style sequential ensemble over neighbor hops. AdaGCN defines layer $l$ as applying a shared two-layer network $f_\theta^{(l)}$ to the precomputed aggregate $\hat{A}^l X$, then uses the soft multi-class AdaBoost variant SAMME.R to turn each round's weighted probability estimates into scores $h_k^{(l)}(x) = (K-1)\left(\log p_k^{(l)}(x) - \frac{1}{K}\sum_{k'} \log p_{k'}^{(l)}(x)\right)$ and to update node weights. The final prediction is $\arg\max_k \sum_{l=0}^L h_k^{(l)}(\hat{A}^l X)$. With this design the paper reports that test accuracy rises as $L$ grows where GCN, residual GCN, and SGC plateau or fall; that AdaGCN posts the best average accuracy on CiteSeer, Cora-ML, PubMed, MS-Academic, and Reddit in its comparison; and that its edge over APPNP widens at lower label rates. It proves in Proposition 2 that AdaGCN can represent general layer-wise neighborhood mixing while stacked GCNs cannot, and argues in Proposition 1 that with shared parameters the ensemble is an exponential-moving-average form of PPNP/APPNP, making AdaGCN their adaptive counterpart.

Load-bearing premise

The theoretical connection to personalized-PageRank models is exact only when the base classifier is linear, yet the accuracy gains the paper reports are attributed to that classifier being nonlinear.

Editorial extensions

If this is right

Adding a round to AdaGCN should improve or hold node-classification accuracy where stacked GCNs oversmooth; the paper shows accuracy increasing with layers on CiteSeer, Cora-ML, and PubMed while GCN and residual GCN decline.
On graphs with very few labels per class, the adaptive combination should help most, since the paper's largest gains over APPNP occur at low label rates (for example, 83.7 versus 80.1 percent on Cora-ML at a 2 percent label rate).
Per-epoch training cost should stay nearly constant with depth because the sparse adjacency matrix is used only to precompute $\hat{A}^l X$; on Reddit the paper reports 32 ms per epoch versus about 29.5 s for APPNP.
The architecture should inherit AdaBoost's generalization behavior, including a VC-dimension-based overfitting warning at very large $L$, which gives a principled reason to choose depth by cross-validation.
AdaGCN should be understood as an adaptive APPNP: fixing its per-layer weights to the exponential decay $(1-\gamma)^l$ and sharing parameters recovers the propagation-based model, so any accuracy gain over APPNP is attributable to adaptivity and per-layer parameters.

Reading between the lines

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

If the mechanism is as general as it looks, the same 'precompute graph powers, then boost a dense classifier' pattern should transfer to other base models, including attention-based or message-passing layers, by feeding them $\hat{A}^l X$; the paper's own claims do not test this.
A natural stress test the paper does not run is heterophily: on graphs where distant neighbors are less informative, the adaptive per-round weights might learn to down-weight unhelpful hops, which would make AdaGCN a candidate replacement for fixed propagation in non-homophilic settings.
The paper's computational comparison suggests a broader design principle for graph models: separate the one-time, sparse aggregation phase from the dense, trainable classifier phase; that principle could be evaluated independently by taking an existing GNN and moving all adjacency multiplies outside the training loop.
Since the exact APPNP equivalence requires a linear classifier, an intermediate architecture with an 'almost linear' $f_\theta$ might preserve the theoretical link while keeping some nonlinearity; the paper does not explore this trade-off.

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

AdaGCN is a genuinely new combination, and the empirical work is more careful than most graph papers. The recipe — one shared MLP applied to successive powers of the normalized adjacency matrix, with training-node weights updated by SAMME.R and the per-hop outputs combined by their AdaBoost coefficients — is a real departure from JK, MixHop, DAGNN, and APPNP. The layer-wise training with shared initialization is a sensible RNN-like framing, not an afterthought. Code is available, and the experiments use 100 runs per setting plus multiple label rates, which is more rigorous than the field average. That deserves credit.

The soft spots are real but not fatal. First, 'consistent state-of-the-art' is over-claimed: in Table 2, the published APPNP number on MS Academic is 93.27 against AdaGCN's 93.17, and the paired t-test p-value is against their own re-implemented APPNP, not the published baseline. Second, the computational-advantage claim is built on per-epoch training time only. The paper itself concedes that 'more epochs are normally needed in the entire layer-wise training,' and the Appendix shows AdaGCN runs 20,000 epochs on Reddit versus 500 for APPNP, with hidden layers orders of magnitude larger. Without total wall-clock time, the 'huge computational efficiency' claim is unsubstantiated. Third, the theoretical connection in Proposition 1 is loose: the proof gets from PPNP to an EMA form by commuting A^l with f_theta, which is only valid for a linear f_theta — the very nonlinearity the paper credits for its representation power. The proposition may work as intuition, but it is not an equivalence. Finally, there is no head-to-head table against DAGNN or MixHop, the closest architectural relatives.

None of this sinks the paper. The core idea is sensible, and the accuracy gains on Cora-ML, CiteSeer, and PubMed are consistently positive and sizable, especially at low label rates. The gaps are addressable in a revision: report end-to-end times, add the missing baselines, and qualify the SOTA statement. For a practitioner, AdaGCN is easy to implement from released code, and the adaptive weighting does seem to help. If I were refereeing this line of work, I would send it to review; the idea deserves referee time even though the current write-up overstates two of its claims. For our own work, I'd cite it as a distinct multi-hop ensemble approach, and I'd avoid repeating the per-epoch timing comparison without a wall-clock check.

Referee Report

3 major / 4 minor

Summary. The paper proposes AdaGCN, a graph neural network architecture for semi-supervised node classification. Instead of stacking graph convolution layers, AdaGCN repeatedly multiplies the input feature matrix by the normalized adjacency matrix to form A^l X, feeds each such matrix to a shared two-layer fully connected network, trains these base classifiers sequentially with the SAMME.R multi-class boosting algorithm, and combines their predictions with boosting-derived weights. The authors claim that this design extracts information from high-order neighbors, mitigates oversmoothing, reaches consistent state-of-the-art accuracy across citation and social graphs at different label rates, and enjoys a large computational advantage over propagation-based methods such as PPNP and APPNP. They also state theoretical connections between AdaGCN and PPNP/APPNP (Proposition 1) and between AdaGCN and MixHop-style neighborhood mixing (Proposition 2), with proofs in the appendix.

Significance. If the empirical and theoretical claims are correct, AdaGCN is a practically useful architecture: it is simple, avoids sparse tensor operations inside the neural network, is backed by a released implementation, and is evaluated with 100 runs per setting and multiple label rates, which is more careful than much of the GNN literature at its time. The paper also attempts to ground the design in boosting theory, which is a potentially valuable direction. However, the strength of the contribution depends on two load-bearing points that the manuscript does not fully establish: the claimed theoretical relation to PPNP/APPNP rests on an invalid step for nonlinear classifiers, and the computational-advantage claim is supported only by per-epoch measurements rather than end-to-end cost. The accuracy results are strong on most datasets but not uniformly state-of-the-art, as the published APPNP baseline exceeds AdaGCN on one dataset.

major comments (3)

[Appendix A.3 and Proposition 1]
[Section 4.3, Table 4, Appendix A.6]
[Table 2]

minor comments (4)

[Abstract and Section 1]
[Table 1 and Table 3]
[Section 5 and Appendix A.5]
[Section 2.2 and Appendix A.3]

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the AdaBoost weights are computed from training error, and the APPNP connection is an illustrative analogy that is not used to fit the reported accuracies.

full rationale

The core combination in AdaGCN is not circular: the coefficients alpha^(l) are computed by Eq. (4) from the weighted error rate of the current base classifier on the training nodes, and Eq. (6) sums the same classifiers with those coefficients, which is exactly the standard AdaBoost/SAMME.R loop rather than a fit of a target quantity. The connection to PPNP/APPNP (Proposition 1) is presented as theory, but the proof's step "put Ahat^l into f_theta" commutes Ahat^l with a nonlinear function, so the claimed equivalence is only heuristic; importantly, this proposition is not used to derive or fit the accuracy numbers in Tables 2-4, so it is not a circular input to the empirical claims. The only self-citation, "Inspired by the Layer Effect on graphs (Sun et al., 2019)" (Sec. 4.2), is interpretive motivation and is not load-bearing because the experimental results are benchmarked against external baselines (APPNP, GCN, GAT, SGC, JK). The paper does concede in Sec. 4.3 that "more epochs are normally needed in the entire layer-wise training" and reports no total wall-clock time, so the computational-advantage claim is under-evidenced, but this is an evidence gap rather than a circular derivation. Overall, no prediction or theoretical claim reduces to its own input by construction; the modest score reflects only the minor, non-load-bearing self-citation.

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

The central claim rests on standard graph-convolution machinery (normalized adjacency propagation), a two-layer MLP base classifier whose capacity is argued rather than proven, and the beta-mixing assumption used to import boosting consistency. Hyperparameters (layer count, hidden size, dropout, weight decay) are tuned per dataset by line search and materially affect the reported accuracy; they are best viewed as fitting choices rather than theory-derived constants. No new physical or mathematical entities are introduced.

free parameters (4)

number of layers L = 15, 12, 20, 5, 4 (Citeseer, Cora-ML, PubMed, MS Academic, Reddit)
Chosen by line search per dataset (Section 4, basic setting). The layer depth strongly affects accuracy and the overfitting behavior described in Appendix A.7.
hidden units h = 5000 (3000 for MS Academic)
Chosen by line search; unusually large hidden layer is a key capacity choice for each base classifier.
dropout rate = 0 for Citeseer/Cora-ML, 0.2 for others
Hand-set per dataset (Section 4 basic setting).
weight decay = 1e-3 for Citeseer, 1e-4 for others
Hand-set per dataset (Section 4 basic setting).

assumptions (4)

domain assumption Normalized adjacency with self-loops, A_hat = D^{-1/2}(A+I)D^{-1/2}, propagates features in the intended way.
Used throughout Eq. 1-3 and Algorithm 1 as the propagation operator; standard in GCN literature.
domain assumption A two-layer fully-connected network with ReLU is a sufficient base classifier for each hop's features.
Section 2.1 argues sparse node features do not need deep nonlinear transformation; this is a modeling hypothesis, not proven.
domain assumption Node samples are drawn from a beta-mixing (weakly dependent) process, so boosting consistency carries over to graphs.
Appendix A.5 invokes Lozano et al. (2013) for consistency of boosting under beta-mixing; whether real citation/social graphs satisfy it is not tested.
standard math ReLU is homogeneous for nonnegative scalings, enabling the construction in Proposition 2.
Appendix A.4 sets W(0) proportional to identity and relies on ReLU(cx)=c ReLU(x) for c>=0; this requires alpha(l) > 0, which holds under the stated condition on err(l).

how reviews work

0 comments

Cite this review

Pith. "Pith review of AdaGCN: Adaboosting Graph Convolutional Networks into Deep Models." pith.science (2026). https://pith.science/paper/6MO6NX7S

@misc{pith2026190805081,
  author       = {Pith},
  title        = {Pith review of: AdaGCN: Adaboosting Graph Convolutional Networks into Deep Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6MO6NX7S}},
  note         = {Machine review of arXiv:1908.05081}
}

read the original abstract

The design of deep graph models still remains to be investigated and the crucial part is how to explore and exploit the knowledge from different hops of neighbors in an efficient way. In this paper, we propose a novel RNN-like deep graph neural network architecture by incorporating AdaBoost into the computation of network; and the proposed graph convolutional network called AdaGCN~(Adaboosting Graph Convolutional Network) has the ability to efficiently extract knowledge from high-order neighbors of current nodes and then integrates knowledge from different hops of neighbors into the network in an Adaboost way. Different from other graph neural networks that directly stack many graph convolution layers, AdaGCN shares the same base neural network architecture among all ``layers'' and is recursively optimized, which is similar to an RNN. Besides, We also theoretically established the connection between AdaGCN and existing graph convolutional methods, presenting the benefits of our proposal. Finally, extensive experiments demonstrate the consistent state-of-the-art prediction performance on graphs across different label rates and the computational advantage of our approach AdaGCN~\footnote{Code is available at \url{https://github.com/datake/AdaGCN}.}

Figures

Figures reproduced from arXiv: 1908.05081 by the authors.

**Figure 2.** Comparison of the graph model architectures. fa in JK network denotes one aggregation layer with aggregation function such as concatenation or max pooling. Architectural Difference. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Comparison of test accuracy of different models as the layer increases. We regard the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Left: Per-epoch training time of AdaGCN vs other methods under 5 runs on four datasets. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: AdaSGC vs AdaGCN. A.3 PROOF OF PROPOSITION 1 Firstly, we further elaborate the Proposition 1 as follows, then we provide the proof. Suppose that γ is the teleport factor. Consider the output ZPPNP = γ(I − (1 − γ)Aˆ) −1 fθ(X) in PPNP and ZAPPNP from its approxminated ve…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 19 canonical work pages

[1]

N-gcn: Multi-scale graph con- volution for semi-supervised node classiﬁcation.International Workshop on Mining and Learning with Graphs (MLG), 2018a

Sami Abu-El-Haija, Amol Kapoor, Bryan Perozzi, and Joonseok Lee. N-gcn: Multi-scale graph con- volution for semi-supervised node classiﬁcation.International Workshop on Mining and Learning with Graphs (MLG), 2018a. Sami Abu-El-Haija, Bryan Perozzi, Amol Kapoor, Hrayr Harutyunyan, Nazanin Alipourfard, Kristina Lerman, Greg Ver Steeg, and Aram Galstyan. Mix...

work page 2019
[3]

Spectral networks and locally connected networks on graphs

Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs. International Conference on Learning Representations, ICLR 2015,

work page 2015
[6]

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., volume 2, pp. 729–734. IEEE,

work page 2005
[11]

Lanczosnet: Multi-scale deep graph convolutional networks

Renjie Liao, Zhizhen Zhao, Raquel Urtasun, and Richard S Zemel. Lanczosnet: Multi-scale deep graph convolutional networks. International Conference on Learning Representations, ICLR 2019,

work page 2019
[13]

Graph attention networks

11 Published as a conference paper at ICLR 2021 Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. International Conference on Learning Representations, ICLR 2019,

work page 2021
[14]

Simplifying graph convolutional networks

Felix Wu, Tianyi Zhang, Amauri Holanda de Souza Jr, Christopher Fifty, Tao Yu, and Kilian Q Weinberger. Simplifying graph convolutional networks. International Conference on Machine Learning, ICML 2019,

work page 2019
[15]

How powerful are graph neural networks? ICLR, 2018a

Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? ICLR, 2018a. Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken-ichi Kawarabayashi, and Stefanie Jegelka. Representation learning on graphs with jumping knowledge networks. International Conference on Machine Learning, ICML 2018, 2018b. Hanqing Zeng...

work page 2018
[17]

12 Published as a conference paper at ICLR 2021 A A PPENDIX A.1 R ELATED WORKS ON DEEP GRAPH MODELS A straightforward solution (Kipf & Welling, 2017; Xu et al., 2018b) inspired by ResNets (He et al.,

work page 2021

Show all 21 references

[18]

utilized Lanczos algorithm to construct low rank approximations of the graph Laplacian and then can exploit multi-scale information. Moreover, APPNP (Approximate Per- sonalized Propagation of Neural Predictions, (Klicpera et al., 2018)) leveraged the relationship between GCN a...

2018
[19]

are also helpful to construct deep graph models. A.2 I NSUFFICIENT REPRESENTATION POWER OF ADASGC As illustrated in Figure 5, with the increasing of layers, AdaSGC with only linear transformation has insufﬁcient representation power both in extracting knowledge from high-order...

2021
[20]

It is strongly Bayes-risk consistent if limn→∞L(hn) = a almost surely

(Consistency) A classiﬁcation rule is consistent for a certain distribution P if E(L(hn)) = P{hn(X) = Y} → a as n → ∞ where a is a constant. It is strongly Bayes-risk consistent if limn→∞L(hn) = a almost surely. Under these deﬁnitions, the convergence and consistence of regula...

2013
[21]

for fair comparison. Furthermore, boosting theory also has the capacity to perfectly incorporate early stopping and it has been shown that for several boosting algorithms including AdaBoost, this regulariza- tion via early stopping can provide guarantees of consistency (Zhang ...

2005
[2003]

Fastgcn: fast learning with graph convolutional networks via importance sampling

Jie Chen, Tengfei Ma, and Cao Xiao. Fastgcn: fast learning with graph convolutional networks via importance sampling. International Conference on Learning Representations, ICLR 2019,

2019
[2004]

Power up! robust graph convolutional network against evasion attacks based on graph powering

10 Published as a conference paper at ICLR 2021 Ming Jin, Heng Chang, Wenwu Zhu, and Somayeh Sojoudi. Power up! robust graph convolutional network against evasion attacks based on graph powering. arXiv preprint arXiv:1905.10029 ,

2021 arXiv
[2005]

Max-margin nonparametric latent feature models for link prediction

Jun Zhu, Jiaming Song, and Bei Chen. Max-margin nonparametric latent feature models for link prediction. arXiv preprint arXiv:1602.07428,

arXiv
[2008]

Pitfalls of graph neural network evaluation.In Relational Representation Learning Workshop (R2L 2018), NeurIPS,

Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan G¨unnemann. Pitfalls of graph neural network evaluation.In Relational Representation Learning Workshop (R2L 2018), NeurIPS,

2018
[2014]

Bootstrapped graph diffusions: Exposing the power of nonlinear- ity

Eliav Buchnik and Edith Cohen. Bootstrapped graph diffusions: Exposing the power of nonlinear- ity. In Abstracts of the 2018 ACM International Conference on Measurement and Modeling of Computer Systems, pp. 8–10. ACM,

2018
[2016]

Learning deep resnet blocks sequentially using boosting theory

Furong Huang, Jordan Ash, John Langford, and Robert Schapire. Learning deep resnet blocks sequentially using boosting theory. International Conference on Machine Learning, ICML 2018,

2018
[2017]

Predict then propagate: Graph neural networks meet personalized pagerank

Johannes Klicpera, Aleksandar Bojchevski, and Stephan G ¨unnemann. Predict then propagate: Graph neural networks meet personalized pagerank. International Conference on Learning Rep- resentations, ICLR 2019,

2019
[2018]

Deep gaussian embedding of graphs: Unsuper- vised inductive learning via ranking

Aleksandar Bojchevski and Stephan G ¨unnemann. Deep gaussian embedding of graphs: Unsuper- vised inductive learning via ranking. International Conference on Learning Representations, ICLR 2019,

2019
[2019]

Semi-supervised classiﬁcation with graph convolutional net- works

Thomas N Kipf and Max Welling. Semi-supervised classiﬁcation with graph convolutional net- works. International Conference on Learning Representations, ICLR 2018,

2018

Pith tools

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

[1] [1]

N-gcn: Multi-scale graph con- volution for semi-supervised node classiﬁcation.International Workshop on Mining and Learning with Graphs (MLG), 2018a

Sami Abu-El-Haija, Amol Kapoor, Bryan Perozzi, and Joonseok Lee. N-gcn: Multi-scale graph con- volution for semi-supervised node classiﬁcation.International Workshop on Mining and Learning with Graphs (MLG), 2018a. Sami Abu-El-Haija, Bryan Perozzi, Amol Kapoor, Hrayr Harutyunyan, Nazanin Alipourfard, Kristina Lerman, Greg Ver Steeg, and Aram Galstyan. Mix...

work page 2019

[2] [3]

Spectral networks and locally connected networks on graphs

Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs. International Conference on Learning Representations, ICLR 2015,

work page 2015

[3] [6]

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., volume 2, pp. 729–734. IEEE,

work page 2005

[4] [11]

Lanczosnet: Multi-scale deep graph convolutional networks

Renjie Liao, Zhizhen Zhao, Raquel Urtasun, and Richard S Zemel. Lanczosnet: Multi-scale deep graph convolutional networks. International Conference on Learning Representations, ICLR 2019,

work page 2019

[5] [13]

Graph attention networks

11 Published as a conference paper at ICLR 2021 Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. International Conference on Learning Representations, ICLR 2019,

work page 2021

[6] [14]

Simplifying graph convolutional networks

Felix Wu, Tianyi Zhang, Amauri Holanda de Souza Jr, Christopher Fifty, Tao Yu, and Kilian Q Weinberger. Simplifying graph convolutional networks. International Conference on Machine Learning, ICML 2019,

work page 2019

[7] [15]

How powerful are graph neural networks? ICLR, 2018a

Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? ICLR, 2018a. Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken-ichi Kawarabayashi, and Stefanie Jegelka. Representation learning on graphs with jumping knowledge networks. International Conference on Machine Learning, ICML 2018, 2018b. Hanqing Zeng...

work page 2018

[8] [17]

12 Published as a conference paper at ICLR 2021 A A PPENDIX A.1 R ELATED WORKS ON DEEP GRAPH MODELS A straightforward solution (Kipf & Welling, 2017; Xu et al., 2018b) inspired by ResNets (He et al.,

work page 2021