Pith. sign in

REVIEW 4 major objections 5 minor 61 references

Boundary Embedding Shaping with Adaptive Contrastive Learning for Graph Structural Disentanglement

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read The paper claims boundary-focused contrastive learning provably disentangles invariant class structure from spurious neighbor correlations, and the plug-in lifts node classification by ~3.3% on average (5.0% on WikiCS).

desk verdict Solid empirical recipe for boundary-focused graph contrastive learning; the attached identifiability theory doesn't cover the loss the method actually trains. read the letter →

arxiv 2606.20283 v2 pith:XW6TFOSY submitted 2026-06-18 cs.LG cs.AI

Jiaqing Chen , Zidu Yin , Yichao Cai , Yuhang Liu , Zhen Zhang , Dong Gong , Javen Qinfeng Shi This is my paper
classification cs.LGcs.AI MSC 68T0768R10
keywords graphstructuralentanglementboundarynodescontrastivelearningneuralnetworksnodeclassificationlatentvariablemodelidentifiabilitylinkprediction
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

This paper argues that the main bottleneck in GNN node classification is not uniform structural noise but structural entanglement concentrated near class boundaries, where irrelevant neighbor information blurs decisions. It introduces Boundary Embedding Shaping (BES), a plug-in module that identifies boundary nodes, pulls their embeddings toward their own class centroid, and pushes them away from other class centroids, while scaling parameter updates by the global embedding perturbation they cause. The paper further claims a theoretical guarantee: under a latent-variable model in which each node is generated from invariant class-determining variables plus variant noise variables, minimizing this boundary-focused contrastive loss block-identifies the invariant variables. If correct, BES is a general, low-parameter add-on that sharpens decision boundaries in existing GNNs and consistently improves node classification and link prediction. A sympathetic reader should care because the method targets the specific failure mode — blurred decision boundaries — rather than treating every node uniformly.

What carries the argument

BES is built on the 'gravity loss,' a contrastive objective defined only for selected boundary nodes. A boundary region B is a covariance-scaled slab around the separating direction between class centroids; within it, nodes whose embedding-space neighbors disagree with their label more than half the time are boundary nodes. For those nodes, the positive similarity is squared distance to the class centroid and the negative similarity is squared distance to other class centroids. The load-bearing identity is Equation (8): the sum of squared distances to all same-class members decomposes into a centroid-distance term plus a constant, so optimizing against the centroid is gradient-equivalent to

What would settle it

Add controlled label noise to a fraction of training nodes on Cora or WikiCS while keeping the graph fixed; BES's claimed gain over GCN should shrink monotonically as the noise fraction grows if positive-pair disentanglement, rather than generic centroid fitting, is responsible for the improvement.

Watch

Extended reading notes

Core claim

At the paper's center is a latent-variable model in which observed node features x = g(z_inv, z_var), with the label y determined solely by z_inv. The main theoretical result, Theorem 3.1, says that minimizing a standard contrastive loss over positive pairs that share z_inv — operationalized as nodes with the same class label — block-identifies z_inv: the learned representation contains all and only invariant information, provided positives vary sufficiently in their variant variables (Assumptions 3.1 and 3.2). Corollary 3.1 restates this as mutual-information disentanglement, and Proposition 3.1 bounds representation error by the optimization loss plus an inverse-margin term. BES operationa

Load-bearing premise

The load-bearing premise is that all nodes sharing a class label have identical invariant latent variables and that the label depends only on those variables, with a practical side condition that class centroids are not highly overlapping in embedding space — exactly the regime BES targets.

Editorial extensions

If this is right

  • BES can be attached to existing GNN encoders as a plug-in and is claimed to consistently improve node classification and link prediction across homogeneous, heterophilic, and OGB benchmark graphs.
  • Under the stated assumptions, the gravity loss removes variant structural information from learned embeddings, so the reported gains are attributed to disentanglement rather than generic centroid fitting.
  • Because only a small attention layer is trained and updates are scaled by the global perturbation they cause, non-boundary nodes are largely preserved while decision boundaries are sharpened.
  • The center-based approximation is claimed to inherit the theoretical guarantees of the full pairwise contrastive loss while reducing complexity from O(N^2) to O(N), enabling large-graph use.
  • Cleaner node embeddings propagate to edge-level tasks: link prediction improves because edge representations inherit less spurious structural noise.

Reading between the lines

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

  • If the identifiability mechanism is doing the work, BES should be most valuable under heterophily, where same-class nodes have strongly varying local neighborhoods; varying the homophily ratio of a synthetic graph and tracking BES's gain over GCN would test this directly.
  • The method's dependence on label-defined positives suggests a stress test: introducing controlled label noise should erode the gain monotonically, because positive pairs would no longer share the same invariant variables.
  • The same boundary-selection and adaptive-update machinery could be ported to low-label or self-supervised settings by replacing class labels with cluster assignments, though the block-identifiability guarantee would then need re-examination.
  • The inverse-margin bound implies that the boundary width parameter could be annealed during training, concentrating updates on the hardest nodes exactly as the margin shrinks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Boundary Embedding Shaping (BES), a GNN plug-in that identifies boundary nodes in embedding space, then applies an adaptive contrastive/centroid-attraction loss to sharpen those nodes' embeddings while leaving non-boundary nodes largely untouched. The authors claim three contributions: (1) a theoretical analysis showing that boundary-focused contrastive learning provably disentangles invariant latent factors from spurious structural correlations; (2) the BES framework combining hard example mining, a center-based gravity loss, and adaptive update scaling; (3) extensive experiments on homogeneous, heterophilic, heterogeneous, and OGB graphs showing consistent gains over strong baselines, e.g., a 3.3% average GCN improvement. The empirical core appears plausible: the paper's own ablations (Tables 14 and 16) show consistent gains from the boundary-aware shaping, and code is provided. However, the theoretical guarantee is not connected to the loss that is actually optimized, the proof of Proposition 3.1 is mathematically unsound as written, and key operational details about label usage in boundary selection are under-specified.

Significance. If the empirical improvements are robust, BES is a useful and simple plug-in: it is architecture-agnostic, adds few parameters, and the controlled ablations indicate that the boundary-selection and centroid-attraction mechanism, rather than the extra encoder, is responsible for the gains. The paper ships code, reports honest w/o-boundary controls, and covers a broad range of datasets. At the same time, the advertised 'provable disentanglement' guarantee is not established for the implemented method: Theorem 3.1 concerns a pairwise InfoNCE objective, while the method trains a centroid-based gravity loss, and Proposition 3.1's error bound does not follow from its proof. These are load-bearing for Contribution 1, not merely presentation issues. With a corrected or removed theory section and recalibrated claims, the empirical contribution is publishable; in its current form the central theoretical claim is unsupported.

major comments (4)
  1. [§4.2 and Theorem 3.1] Thm. 3.1 guarantees block-identifiability only for the pairwise InfoNCE loss with a general similarity. The loss actually optimized (Eqs. 9–10, Algorithm 1) replaces positive similarities with a single class-centroid distance and negatives with other class centroids. The identity in Eq. (8) justifies only the positive-sum reduction: the second term is constant w.r.t. Φ(x_b), so it does not affect the gradient. It does not justify replacing the denominator's sum of exponentials over negative similarities, because Σ exp(-||Φ(x_b)-Φ(x_n)||²) ≠ exp(-||Φ(x_b)-μ_j||²) in general. Hence the 'gradient-exact reduction' stated at the end of §4.2 does not hold for the full contrastive objective, and the reported results do not train the loss for which identifiability was proved. Appendix D.2 confirms that the exact pairwise loss was abandoned in favor of the center approximation. Please either prov
  2. [Assumption 3.1 / §3.1 and §4.1] Asm. 3.1 requires positive pairs to satisfy z_inv = z̃_inv almost surely. The operational positive-pair construction in §4.1 and Algorithm 1 groups nodes by observed class label: positives are nodes in C_{y_b}. Since Eq. (3) only assumes y = g_y(z_inv) with g_y deterministic but not injective, equal label does not imply equal z_inv. Thus the key premise of Thm. 3.1 is not guaranteed by the method's own sampling rule. The paper provides no experiment (e.g., synthetic data with known latent factors) that checks whether invariant latent variables are actually constant within classes. The 'provably disentangles' claim should be weakened to a conditional statement, or an injectivity/refinement assumption on g_y must be explicitly stated and argued for the benchmarks.
  3. [Appendix A.3, Prop. 3.1 / Eq. (33)] The proof of Prop. 3.1 is incorrect as written. Eq. (33) gives E∥Rhat{z}−z*∥² ≤ 2ε/(c·margin). Taking square roots and applying Jensen's inequality yields E∥Rhat{z}−z*∥ ≤ sqrt(2ε/(c·margin)), which is O(sqrt(ε/margin)), not the claimed O(sqrt ε) + O(1/margin). The final bound in Eq. (34) therefore does not follow. Additionally, the curvature-to-margin relation λ_min ≥ c·margin(∂Z_inv) in Eq. (32) is asserted without proof, and 'margin(∂Z_inv)' is never formally defined. Since Prop. 3.1 is the stated theoretical basis for the adaptive update in §4.3, that motivational chain is currently unsupported. The empirical adaptive mechanism may stand on its own, but the proposition should be fixed, proved, or removed from the theoretical claims.
  4. [Algorithm 1 / §4.1] The boundary-selection procedure is under-specified regarding training/test labels, creating a validity risk. Eq. (5) defines the boundary region B over all xi ∈ X, and Eq. (4) computes S(vi) for vi ∈ B using neighbor labels y_j. The text states that 'identification explicitly leverages label information solely from training nodes, while strictly excluding all test nodes and test labels,' but Algorithm 1 does not show how this restriction is operationalized. If B includes test nodes, Eq. (4) would require test labels; if it does not, the pseudocode should say that B is intersected with the training set before computing S. Please clarify the exact protocol, including how boundary nodes among test nodes are selected (if any) and whether validation/test features or labels are used to estimate μ_m, Σ_m. Without this clarification, the reproducibility and integrity of the empirical comparison
minor comments (5)
  1. [Table 1 and Abstract] Table 1 reports no standard deviations. Given that several margins are small (e.g., Cora NC: 89.46 vs 88.72; Cora LP: 94.45 vs IGCL's 95.13), error bars and significance tests should be reported. The abstract's 'consistently outperforms existing leading methods' is contradicted by the Cora LP row, where IGCL is higher; please qualify the claim.
  2. [§C.4 and Table 3] The hyperparameters specific to BES (δ, α, β, boundary threshold, number of layers L, number of iterations Q) are not listed in Table 3 or in §C.4. The sensitivity analysis in Fig. 3(4) mentions optimal τ=1 and δ=5, but exact per-dataset settings needed for reproduction are missing. Please add a table of BES hyperparameters.
  3. [Eq. (5)] The covariance Σ in the Mahalanobis slab is not defined. State whether it is a pooled covariance, a class-conditional covariance, or a global covariance, and specify how it is estimated (with or without labels). This detail matters because B (and hence the selected boundary nodes) depends on it.
  4. [Algorithm 1] Notation: B denotes both the boundary region from Eq. (5) and the selected boundary-node set in line 8 of Algorithm 1. Please rename one (e.g., B_region and B_nodes) to avoid ambiguity.
  5. [References and terminology] Reference inconsistencies: GCN is cited as (Yang et al., 2016) in Table 1 but the reference list has Kipf & Welling (2017) for GCN. Also, 'heterogeneous' and 'heterophilic' are used interchangeably in §5 and Appendix C; these should be made consistent.

Circularity Check

3 steps flagged · score 6.0 of 10

Identifiability guarantee is built into the label-based positive-pair assumption, and the theorem's objective is not the loss actually optimized.

  1. self definitional [Section 3.1, Eq. (3) and Assumption 3.1; Section 4.1, Eqs. (6)-(7)]
    "z = (z_inv, z_var) ~ p_z, y = g_y(z_inv), x = g(z) ... gy is a deterministic function indicating that z_inv exclusively determines the class label y. Assumption 3.1: ... the two nodes ... share the same class label y. Then, the conditional density p_~z|z(~z|z) = delta(~z_inv - z_inv)p_~zvar|zvar(~zvar|zvar) ... For a designated boundary node, ... the embedding Phi(x_b) should be aligned with all positive samples v_p sharing the same label y_b."

    The variable to be recovered (z_inv) is defined by the model as the class-determining factor, and Assumption 3.1 stipulates that positive pairs sharing a class label have identical z_inv. The method then constructs positive pairs precisely by class label (Section 4.1). Therefore the theorem's conclusion that contrastive learning isolates z_inv is an unpacking of the input assumption: label-grouping is assumed to equal invariant-factor grouping. No derivation or test establishes that same-label nodes share z_inv or that z_var is excluded; the claimed disentanglement from spurious structure is built into the definition of the positives.

  2. self citation load bearing [Appendix A.1, Step 4 (proof of Theorem 3.1)]
    "By directly applying Lemma B.2 in (Cai et al., 2025), we conclude that the learned representations contain all and only the information about the invariant latent variables z_inv almost surely."

    This is the step that upgrades the proof from 'the learned representation depends only on z_inv and is uniform' to the full block-identifiability conclusion 'all and only'. The cited Lemma B.2 comes from Cai et al. (2025), which shares authors with the present paper (Yichao Cai, Yuhang Liu, Zhen Zhang, Javen Qinfeng Shi). The lemma is not proved in this manuscript, is not machine-checked, and no independent external verification is supplied. The 'uniqueness' claim that makes Theorem 3.1's conclusion is thus imported from the authors' own prior work rather than derived from first principles here.

1 more flagged steps
  1. other [Section 4.2, Eq. (8); Appendix D.2]
    "The approximated similarity thus inherits all theoretical properties of the pairwise form, which is a gradient-exact reduction ... We initially attempted to optimize the exact pairwise contrastive loss (Thm. 3.1) ... this yielded extensive computational overhead and chaotic convergence. ... By operating on cluster centers instead of node instances, theoretical properties (identifiability, information bounds) established for individual pairs efficiently remain preserved through geometric identity."

    This is the load-bearing bridge from Theorem 3.1 to the implemented loss: the paper claims the centroid-based gravity loss inherits the theorem's identifiability guarantee by a gradient-exact reduction. But Eq. (8) is an identity only for the sum of squared distances to the class centroid, i.e. the positive numerator. The denominator of the Theorem 3.1 InfoNCE objective contains exp(sim_neg) summed over negative samples; a sum of exponentials is not the exponential of a centroid distance. Appendix D.2 confirms the exact pairwise loss was abandoned. Consequently the trained gravity loss is not the objective for which identifiability was proved, and the claim that the theoretical properties 'remain preserved' is asserted rather than derived for the full contrastive loss.

full rationale

The empirical benchmark tables are standard external comparisons, and those results are not circular in themselves. The circularity is concentrated in the paper's central theoretical contribution. First, the invariant latent variable is defined as the label-determining content (y = g_y(z_inv)), and Assumption 3.1 stipulates that label-sharing positive pairs have identical z_inv; the method then picks positives by label, so the theorem's 'disentanglement' is a restatement of the construction of the positives rather than an empirically grounded discovery about graph structure. Second, the final 'all and only' step of the theorem is delegated to Lemma B.2 of the authors' own prior work, a self-citation that carries the uniqueness conclusion. Third, the objective actually optimized in the algorithm is the centroid gravity loss, while Theorem 3.1 is proved for a pairwise InfoNCE-style loss; the claimed gradient-exact inheritance is only shown for the positive numerator (Eq. 8), not for the nonlinear negative denominator, and Appendix D.2 states the pairwise loss was abandoned. These issues make the central 'provably disentangles' claim partially circular by construction, even though the empirical gains over baselines may be real and are evaluated externally. Score 6 reflects partial circularity in the central claim, not in the experimental comparison.

Assumptions & free parameters 9 free parameters · 8 assumptions · 3 invented entities

The central claim rests on: (i) an imported identifiability framework whose load-bearing assumption (Asm 3.1) — same-label nodes have identical z_inv — is essentially the conclusion the method needs; (ii) an asserted curvature-margin relation (Eq. 32) that manufactures the O(1/margin) term used to justify the adaptive update; (iii) a self-cited, unstated lemma (Cai et al. 2025, B.2) completing Thm 3.1; and (iv) several tuned hyperparameters (delta, alpha, beta, thresholds). The empirical core is the ~1% gain over an equal-architecture baseline, with the rest of the headline effect coming from stronger encoders and decoder.

free parameters (9)
  • delta (boundary slab width) = 5
    Eq. 5 defines the boundary region B via |(mu_m - mu_n)^T Sigma^-1 Phi(x_i)| <= delta; tuned on benchmarks (Fig. 3(4) reports optimal delta=5).
  • alpha (adaptive update scaling) = not specified in text
    Eq. 13 scales the update by alpha/Delta(B); a tuned hyperparameter controlling update magnitude.
  • beta (boundary batch size) = not specified in text
    Algorithm 1 samples beta boundary nodes per iteration; tuned.
  • eta (learning rate) = 0.01
    Eq. 11 virtual pre-update learning rate; set in Table 3.
  • shift score threshold = 0.5
    Nodes with S(v) > 0.5 are designated boundary nodes (Section 4); threshold chosen by hand.
  • number of attention layers L = 2
    Two attention layers A(B)_1, A(B)_2; chosen by hand (Section C.1).
  • number of encoders K = 2 (1 for ogbl-collab)
    Encoder stack width varies per dataset (Table 13); an architecture choice that contributes to gains beyond the shaping module.
  • temperature tau = 1 (reported 'optimal')
    tau appears in Thm 3.1 and in the sensitivity plot (Fig. 3(4)), but not in the implemented loss (Eqs. 9-10); reporting sensitivity for a hyperparameter the method does not use.
  • kNN neighborhood size k = unspecified
    Eq. 4's N_emb(v_i) requires a k for k-nearest neighbors; k is never stated in the paper.
assumptions (8)
  • domain assumption y = g_y(z_inv): class label fully determined by invariant latent variables
    Eq. 3; if labels depend on z_var or are noisy, Asm 3.1 fails and the whole identifiability argument collapses.
  • domain assumption Assumption 3.1: label-informed positive pairs have identical z_inv
    Section 3.1; the method constructs positives by label, so the conclusion is substantially contained in this premise.
  • domain assumption Assumption 3.2: sufficient local variation of z_var around every conditioning point
    Section 3.1; needed for the contradiction step in the proof of Thm 3.1 and never verified on graph data.
  • standard math InfoNCE asymptotic alignment-maxent decomposition and Darmois construction (von Kuegelgen et al. 2021; Daunhawer et al. 2023)
    Step 1 of the proof of Thm 3.1 (Section A.1) imports these results by citation; accepted as background.
  • ad hoc to paper Lemma B.2 of (Cai et al., 2025): final 'all and only' block-identification step
    Section A.1 Step 4 rests on a lemma from the authors' own concurrent paper, not restated here; the theorem is incomplete without it.
  • ad hoc to paper lambda_min >= c * margin(dZ_inv): curvature of loss Hessian proportional to class-boundary margin
    Eq. 32 in the proof of Prop 3.1 is asserted with a heuristic paragraph; it is the source of the O(1/margin) term used to justify the adaptive update.
  • domain assumption Class centroids are not highly overlapping
    Section 4.2 explicitly conditions the gradient-equivalence on non-overlapping centroids; the target regime is precisely overlapping, entangled classes.
  • domain assumption Embedding-space kNN neighborhoods preserve heterophilic structure of the original graph
    Section D.1 asserts empirical equivalence; the shift score and boundary set depend on this transfer.
invented entities (3)
  • Boundary region B (Eq. 5)
    purpose: Restricts contrastive shaping to nodes in a margin slab around the class-center separating direction
    Constructed via tuned delta and Gaussian cluster parameters; support is only the paper's own ablations (Table 16), no external falsifiable handle.
  • Pre-update perturbation Delta(B) (Eq. 12)
    purpose: Measures global embedding change from a virtual update and inversely scales the actual parameter update
    A designed mechanism with no external handle; its justification is the asserted margin-curvature relation of Prop 3.1.
  • Gravity similarity with per-class min offset (Eq. 9)
    purpose: Proxy positive similarity for the contrastive objective
    Differs from the theorem's loss; the inheritance of identifiability properties is claimed, not derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Boundary Embedding Shaping with Adaptive Contrastive Learning for Graph Structural Disentanglement." pith.science (2026). https://pith.science/paper/XW6TFOSY

@misc{pith2026260620283,
  author       = {Pith},
  title        = {Pith review of: Boundary Embedding Shaping with Adaptive Contrastive Learning for Graph Structural Disentanglement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XW6TFOSY}},
  note         = {Machine review of arXiv:2606.20283}
}
read the original abstract

Graph neural networks (GNNs) excel at aggregating neighbor information for classification, yet their performance is hindered by graph structural entanglement, where spurious correlations from semantically irrelevant neighbors contaminate node embeddings. This challenge is most acute for nodes near class boundaries in the embedding space, where amplified structural noise blurs decision boundaries and destabilizes predictions. Existing robust GNN methods largely treat all nodes uniformly, ignoring boundary vulnerabilities. In this paper, to improve classification performance, we tackle graph structural disentanglement by identifying boundary-region entanglement as the primary bottleneck and propose Boundary Embedding Shaping (BES), an adaptive contrastive learning GNN plug-in module that selectively suppresses spurious structural noise at decision boundaries with minimal model parameter perturbation. Extensive experiments demonstrate that BES consistently improves boundary discrimination and outperforms existing leading methods. Notably, BES boosts GCN performance by an average of 3.3% in node classification (up to 5.0% on WikiCS) and achieves superior accuracy in link prediction.

Figures

Figures reproduced from arXiv: 2606.20283 by the authors.

Figure 1
Figure 1. Boundary embedding shaping on WikiCS. BES structurally disentangles boundary nodes between clusters, producing sharply separable embeddings that significantly improve classification performance. hard example mining (Shrivastava et al., 2016; Xia et al., 2022) to identify boundary nodes, namely nodes predomi￾nantly surrounded by neighbors from different classes. Sub￾sequently, we selectively improve only those nodes … view at source ↗
Figure 2
Figure 2. Boundary embedding shaping architecture aims to filter the irrelevant noise introduced by the variability of the local neighboring topology to obtain improved node embeddings for better classification. k-nearest neighbors), and yi is the category label of node vi . Instead of evaluating shift scores globally across the graph, we first coarsely mark the boundary region B between clusters m and n using a margin-based … view at source ↗
Figure 3
Figure 3. Analysis of boundary embedding shaping. calized noise and exhibit high computational costs on large graphs. Contrastive Representation Learning. Class-guided contrastive learning has become a cornerstone of graph rep￾resentation learning, with seminal works such as Supervised Contrastive Learning (SCL) (Khosla et al., 2020), Proto￾typical Contrastive Learning (PCL) (Li et al., 2021), and GraphECL (Xiao et al., 2024)… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: T-SNE visualization of node embeddings generated by different graph encoders and their combinations [PITH_FULL_IMAGE:figures/full_fig_p023_4.png]
Figure 5
Figure 5. Figure 5: Boundary embedding shaping. (a) is for the binary scenario; (b) is for the multi-class scenario. (a) Raw Embedding (b) Boundary Embedding Shaping (c) Final Node Embeddings [PITH_FULL_IMAGE:figures/full_fig_p026_5.png]
Figure 6
Figure 6. Figure 6: Boundary embedding shaping for Cora. (a) Raw Embedding (b) Boundary Embedding Shaping (c) Final Node Embeddings [PITH_FULL_IMAGE:figures/full_fig_p026_6.png]
Figure 7
Figure 7. Figure 7: Boundary embedding shaping for CiteSeer [PITH_FULL_IMAGE:figures/full_fig_p026_7.png]
Figure 8
Figure 8. Figure 8: Metric convergence of node classification on different datasets during training. (a) Cora (b) CiteSeer (c) Pubmed (d) WikiCS [PITH_FULL_IMAGE:figures/full_fig_p027_8.png]
Figure 9
Figure 9. Figure 9: Metric convergence of link prediction on different datasets during training. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_9.png]

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 2 linked inside Pith

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Z., van den Hengel, A., and Qi, Y

    Bao, S., Xue, Z., Chen, Q., Ou, S., Beheshti, A., Sheng, Q. Z., van den Hengel, A., and Qi, Y. Causal MVC : Causal content-style representation learning for deep multi-view clustering. In Proceedings of the International Conference on Multimedia (MM), pp.\ 1598--1606, 2025

    2025

  3. [3]

    Bishop, C. M. Latent variable models. In Learning in Graphical Models, volume 89, pp.\ 371--403. Springer, 1998

    1998

  4. [4]

    How attentive are graph attention networks? In Proceedings of the International Conference on Learning Representations (ICLR), 2022

    Brody, S., Alon, U., and Yahav, E. How attentive are graph attention networks? In Proceedings of the International Conference on Learning Representations (ICLR), 2022

    2022

  5. [5]

    Cai, Y., Liu, Y., Gao, E., Jiang, T., Zhang, Z., van den Hengel, A., and Shi, J. Q. On the value of cross-modal misalignment in multimodal representation learning. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), 2025

    2025

  6. [6]

    Nodeimport: Imbalanced node classification with node importance assessment

    Chen, N., Liu, Z., Hooi, B., He, B., Hu, J., and Chen, J. Nodeimport: Imbalanced node classification with node importance assessment. In Proceedings of the ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), pp.\ 94--105, 2025 a

    2025

  7. [7]

    Chen, T., Kornblith, S., Norouzi, M., and Hinton, G. E. A simple framework for contrastive learning of visual representations. In Proceedings of the International Conference on Machine Learning (ICML), pp.\ 1597--1607, 2020

    2020

  8. [8]

    Improving graph contrastive learning with community structure

    Chen, X., Yue, K., Duan, L., and Yu, L. Improving graph contrastive learning with community structure. In Proceedings of the Conference on Uncertainty in Artificial Intelligence (UAI), pp.\ 568--585, 2025 b

    2025

Show all 61 references
  1. [9]

    Daunhawer, I., Bizeul, A., Palumbo, E., Marx, A., and Vogt, J. E. Identifiability results for multimodal contrastive learning. In Proceedings of the International Conference on Learning Representations (ICLR), 2023

    2023

  2. [10]

    Convolutional neural networks on graphs with fast localized spectral filtering

    Defferrard, M., Bresson, X., and Vandergheynst, P. Convolutional neural networks on graphs with fast localized spectral filtering. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), pp.\ 3837--3845, 2016

    2016

  3. [11]

    Imagenet: A large-scale hierarchical image database

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 248--255, 2009

    2009

  4. [12]

    The mnist database of handwritten digit images for machine learning research

    Deng, L. The mnist database of handwritten digit images for machine learning research. IEEE Signal Processing Magazine, 29 0 (6): 0 141--142, 2012

    2012

  5. [13]

    L., Bollacker, K

    Giles, C. L., Bollacker, K. D., and Lawrence, S. Citeseer: An automatic citation indexing system. In Proceedings of the ACM International Conference on Digital Libraries (DL), pp.\ 89--98, 1998

    1998

  6. [14]

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

    Hamilton, W. L., Ying, Z., and Leskovec, J. Inductive representation learning on large graphs. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), pp.\ 1024--1034, 2017

    2017

  7. [15]

    Hoang, V. T. and Lee, O. Transitivity-preserving graph representation learning for bridging local connectivity and role-based similarity. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pp.\ 12456--12465, 2024

    2024

  8. [16]

    Open graph benchmark: Datasets for machine learning on graphs

    Hu, W., Fey, M., Zitnik, M., Dong, Y., Ren, H., Liu, B., Catasta, M., and Leskovec, J. Open graph benchmark: Datasets for machine learning on graphs. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), 2020

    2020

  9. [17]

    Supervised contrastive learning

    Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., and Krishnan, D. Supervised contrastive learning. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), 2020

    2020

  10. [18]

    Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks. In Proceedings of the International Conference on Learning Representations (ICLR), 2017

    2017

  11. [19]

    Koh, P. W. and Liang, P. Understanding black-box predictions via influence functions. In Proceedings of the International Conference on Machine Learning (ICML), pp.\ 1885--1894, 2017

    2017

  12. [20]

    Li, J., Zhou, P., Xiong, C., and Hoi, S. C. Prototypical contrastive learning of unsupervised representations. In Proceedings of the International Conference on Learning Representations (ICLR), 2021

    2021

  13. [21]

    Kolmogorov--arnold graph neural networks for molecular property prediction

    Li, L., Zhang, Y., Wang, G., and Xia, K. Kolmogorov--arnold graph neural networks for molecular property prediction. Nature Machine Intelligence, 7 0 (8): 0 1346--1354, 2025

    2025

  14. [22]

    Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., van den Hengel, A., Zhang, K., and Shi, J. Q. Identifiable latent polynomial causal models through the lens of change. In Proceedings of the International Conference on Learning Representations (ICLR), 2024

    2024

  15. [23]

    Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., van den Hengel, A., Zhang, K., and Shi, J. Q. Latent covariate shift: Unlocking partial identifiability for multi-source domain adaptation. Transactions on Machine Learning Research, 2025. ISSN 2835-8856

    2025

  16. [24]

    Liu, Y., Gong, D., Cai, Y., Gao, E., Zhang, Z., Huang, B., Gong, M., van den Hengel, A., and Shi, J. Q. I predict therefore i am: Is next token prediction enough to learn human-interpretable concepts from data? In Proceedings of the International Conference on Learning Represe...

    2026

  17. [25]

    Liu, Y., Zhang, Z., Gong, D., Gao, E., Huang, B., Gong, M., van den Hengel, A., Zhang, K., and Shi, J. Q. Beyond DAG s: A latent partial causal model for multimodal learning. In Proceedings of the International Conference on Learning Representations (ICLR), 2026 b

    2026

  18. [26]

    Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., van den Hengel, A., Zhang, K., and Shi, J. Q. Identifying weight-variant latent causal models. Journal of Machine Learning Research, 27 0 (4): 0 1--49, 2026 c

    2026

  19. [27]

    Classic gnns are strong baselines: Reassessing gnns for node classification

    Luo, Y., Shi, L., and Wu, X.-M. Classic gnns are strong baselines: Reassessing gnns for node classification. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), pp.\ 97650--97669, 2024

    2024

  20. [28]

    Enhancing robustness of graph neural networks on social media with explainable inverse reinforcement learning

    Lyu, Y., Li, C., Xie, S., and Zhang, X. Enhancing robustness of graph neural networks on social media with explainable inverse reinforcement learning. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), pp.\ 31736--31758, 2024

    2024

  21. [29]

    Disentangled graph convolutional networks

    Ma, J., Cui, P., Kuang, K., Wang, X., and Zhu, W. Disentangled graph convolutional networks. In Proceedings of the International Conference on Machine Learning (ICML), pp.\ 4212--4221, 2019

    2019

  22. [30]

    K., Nigam, K., Rennie, J., and Seymore, K

    McCallum, A. K., Nigam, K., Rennie, J., and Seymore, K. Automating the construction of internet portals with machine learning. Information Retrieval, 3: 0 127--163, 2000

    2000

  23. [31]

    and Cangea, C

    Mernyei, P. and Cangea, C. Wiki-cs: A wikipedia-based benchmark for graph neural networks. arXiv preprint arXiv:2007.02901, 2020

    2007 arXiv

  24. [32]

    Interpretable and generalizable graph learning via stochastic attention mechanism

    Miao, S., Liu, M., and Li, P. Interpretable and generalizable graph learning via stochastic attention mechanism. In Proceedings of the International Conference on Machine Learning (ICML), pp.\ 15524--15543, 2022

    2022

  25. [33]

    L., Lenssen, J

    Morris, C., Ritzert, M., Fey, M., Hamilton, W. L., Lenssen, J. E., Rattan, G., and Grohe, M. Weisfeiler and leman go neural: Higher-order graph neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pp.\ 4602--4609, 2019

    2019

  26. [34]

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

    Pei, H., Wei, B., Chang, K. C.-C., Lei, Y., and Yang, B. Geom-gcn: Geometric graph convolutional networks. In Proceedings of the International Conference on Learning Representations (ICLR), 2020

    2020

  27. [35]

    Pennington, J., Socher, R., and Manning, C. D. Glove: Global vectors for word representation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), pp.\ 1532--1543, 2014

    2014

  28. [36]

    A critical look at the evaluation of gnns under heterophily: Are we really making progress? In Proceedings of the International Conference on Learning Representations (ICLR), 2023

    Platonov, O., Kuznedelev, D., Diskin, M., Babenko, A., and Prokhorenkova, L. A critical look at the evaluation of gnns under heterophily: Are we really making progress? In Proceedings of the International Conference on Learning Representations (ICLR), 2023

    2023

  29. [37]

    K., Bronstein, M

    Rusch, T. K., Bronstein, M. M., and Mishra, S. A survey on oversmoothing in graph neural networks. arXiv preprint arXiv:2303.10993, 2023

    2023 arXiv

  30. [38]

    Collective classification in network data

    Sen, P., Namata, G., Bilgic, M., Getoor, L., Galligher, B., and Eliassi-Rad, T. Collective classification in network data. AI Magazine, 29 0 (3): 0 93--106, 2008

    2008

  31. [39]

    Training region-based object detectors with online hard example mining

    Shrivastava, A., Gupta, A., and Girshick, R. Training region-based object detectors with online hard example mining. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 761--769, 2016

    2016

  32. [40]

    Networked inequality: Preferential attachment bias in graph neural network link prediction

    Subramonian, A., Sagun, L., and Sun, Y. Networked inequality: Preferential attachment bias in graph neural network link prediction. In Proceedings of the International Conference on Machine Learning (ICML), pp.\ 46891--46925, 2024

    2024

  33. [41]

    Sun, L., Huang, Z., Wang, Z., Wang, F., Peng, H., and Yu, P. S. Motif-aware riemannian graph neural network with generative-contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pp.\ 9044--9052, 2024

    2024

  34. [42]

    Graph attention networks

    Velickovic, P., Cucurull, G., Casanova, A., Romero, A., Li \` o , P., and Bengio, Y. Graph attention networks. In Proceedings of the International Conference on Learning Representations (ICLR), 2018

    2018

  35. [43]

    u gelgen, J., Sharma, Y., Gresele, L., Brendel, W., Sch \

    von K \" u gelgen, J., Sharma, Y., Gresele, L., Brendel, W., Sch \" o lkopf, B., Besserve, M., and Locatello, F. Self-supervised learning with data augmentations provably isolates content from style. In Proceedings of the Conference on Neural Information Processing Systems (Ne...

    2021

  36. [44]

    Generalizing aggregation functions in gnns: Building high capacity and robust gnns via nonlinear aggregation

    Wang, B., Jiang, B., Tang, J., and Luo, B. Generalizing aggregation functions in gnns: Building high capacity and robust gnns via nonlinear aggregation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45 0 (11): 0 13454--13466, 2023

    2023

  37. [45]

    Disentangled graph collaborative filtering

    Wang, X., Jin, H., Zhang, A., He, X., Xu, T., and Chua, T. Disentangled graph collaborative filtering. In Proceedings of the ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), pp.\ 1001--1010, 2020

    2020

  38. [46]

    Simplifying graph convolutional networks

    Wu, F., Souza, A., Zhang, T., Fifty, C., Yu, T., and Weinberger, K. Simplifying graph convolutional networks. In Proceedings of the International Conference on Machine Learning (ICML), pp.\ 6861--6871, 2019

    2019

  39. [47]

    Graph neural networks in recommender systems: A survey

    Wu, S., Sun, F., Zhang, W., Xie, X., and Cui, B. Graph neural networks in recommender systems: A survey. ACM Computing Surveys, 55 0 (5): 0 1--37, 2022

    2022

  40. [48]

    Xia, J., Wu, L., Wang, G., Chen, J., and Li, S. Z. Pro GCL : Rethinking hard negative mining in graph contrastive learning. In Proceedings of the International Conference on Machine Learning (ICML), pp.\ 24332--24346, 2022

    2022

  41. [49]

    C., Wang, S., and Honavar, V

    Xiao, T., Zhu, H., Zhang, Z., Guo, Z., Aggarwal, C. C., Wang, S., and Honavar, V. G. Efficient contrastive learning for fast and accurate inference on graphs. In Proceedings of the International Conference on Machine Learning (ICML), pp.\ 54363--54381, 2024

    2024

  42. [50]

    Revisiting semi-supervised learning with graph embeddings

    Yang, Z., Cohen, W., and Salakhudinov, R. Revisiting semi-supervised learning with graph embeddings. In Proceedings of the International Conference on Machine Learning (ICML), pp.\ 40--48, 2016

    2016

  43. [51]

    v., and Locatello, F

    Yao, D., Xu, D., Lachapelle, S., Magliacane, S., Taslakian, P., Martius, G., K \"u gelgen, J. v., and Locatello, F. Multi-view causal representation learning with partial observability. In Proceedings of the International Conference on Learning Representations (ICLR), 2024

    2024

  44. [52]

    Do transformers really perform badly for graph representation? In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), pp.\ 28877--28888, 2021

    Ying, C., Cai, T., Luo, S., Zheng, S., Ke, G., He, D., Shen, Y., and Liu, T. Do transformers really perform badly for graph representation? In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), pp.\ 28877--28888, 2021

    2021

  45. [53]

    Graph contrastive learning with augmentations

    You, Y., Chen, T., Sui, Y., Chen, T., Wang, Z., and Shen, Y. Graph contrastive learning with augmentations. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), pp.\ 5812--5823, 2020

    2020

  46. [54]

    Zeng, H., Zhou, H., Srivastava, A., Kannan, R., and Prasanna, V. K. Graph SAINT : Graph sampling based inductive learning method. In Proceedings of the International Conference on Learning Representations (ICLR), 2020

    2020

  47. [55]

    Decouple graph neural networks: Train multiple simple gnns simultaneously instead of one

    Zhang, H., Zhu, Y., and Li, X. Decouple graph neural networks: Train multiple simple gnns simultaneously instead of one. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46 0 (11): 0 7451--7462, 2024 a

    2024

  48. [56]

    Graph neural networks in modern ai-aided drug discovery

    Zhang, O., Lin, H., Zhang, X., Wang, X., Wu, Z., Ye, Q., Zhao, W., Wang, J., Ying, K., Kang, Y., et al. Graph neural networks in modern ai-aided drug discovery. Chemical Reviews, 125 0 (20): 0 10001--10103, 2025

    2025

  49. [57]

    Trans GNN : Harnessing the collaborative power of transformers and graph neural networks for recommender systems

    Zhang, P., Yan, Y., Zhang, X., Li, C., Wang, S., Huang, F., and Kim, S. Trans GNN : Harnessing the collaborative power of transformers and graph neural networks for recommender systems. In Proceedings of the ACM SIGIR Conference on Research and Development in Information Retri...

    2024

  50. [58]

    and Koniusz, P

    Zhu, H. and Koniusz, P. Simple spectral graph convolution. In Proceedings of the International Conference on Learning Representations (ICLR), 2021

    2021

  51. [59]

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

    Zhu, J., Yan, Y., Zhao, L., Heimann, M., Akoglu, L., and Koutra, D. Beyond homophily in graph neural networks: Current limitations and effective designs. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS), pp.\ 7793--7804, 2020

    2020

  52. [60]

    Graph contrastive learning with adaptive augmentation

    Zhu, Y., Xu, Y., Yu, F., Liu, Q., Wu, S., and Wang, L. Graph contrastive learning with adaptive augmentation. In Proceedings of the Web Conference (WWW), pp.\ 2069--2080, 2021

  53. [61]

    S., Sharma, Y., Schneider, S., Bethge, M., and Brendel, W

    Zimmermann, R. S., Sharma, Y., Schneider, S., Bethge, M., and Brendel, W. Contrastive learning inverts the data generating process. In Proceedings of the International Conference on Machine Learning (ICML), pp.\ 12979--12990, 2021

    2021

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.