Pith. sign in

REVIEW 5 major objections 6 minor 56 references

Graph Community Augmentation with GMM-based Modeling in Latent Space

T0 review · 5 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read This paper proposes graph community augmentation (GCA), a latent-space Gaussian mixture method for generating graphs with a new community while preserving the original graph's structure.

desk verdict Genuinely new framing for graph community augmentation, but the headline claim that the added latent cluster decodes to a real graph community is never directly measured. read the letter →

arxiv 2412.01163 v1 pith:VPFROKJU submitted 2024-12-02 cs.LG cs.ITmath.ITstat.ML

Shintaro Fukushima , Kenji Yamanishi This is my paper · ORCID
classification cs.LGcs.ITmath.ITstat.ML
keywords graphcommunityaugmentationgenerationGaussianmixturemodellatentspaceminimumdescriptionlengthvariationalautoencoderknowledgeextrapolationdetection
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

The paper proposes graph community augmentation (GCA): generating graphs that contain a community absent from the training graph, while keeping the original graph's structure realistic enough that the new graph still looks like it belongs to the same family. The authors' claim is that this reduces to editing the latent density learned by a variational graph autoencoder: fit a Gaussian mixture model to the node embeddings, then add one new Gaussian component whose location and weight are chosen so that it is far from existing clusters (novelty) yet the whole mixture stays close to the original density (reliability), with both constraints expressed through KL divergence and the cluster count selected by the MDL principle. They demonstrate on synthetic stochastic block model graphs and on Cora, PubMed, CiteSeer, and Coauthor-Physics that the generated graphs score higher on anomaly and novelty measures than diffusion- and RNN-based generators while retaining comparable degree, clustering, orbit, and spectral statistics. If correct, GCA would give a principled way to extrapolate graph structure, with potential uses in discovering emerging purchaser communities in social networks and in augmenting scarce real graph data for downstream models.

What carries the argument

The load-bearing object is the Gaussian mixture model in the latent space of a variational graph autoencoder (VGAE). Each component is taken to correspond to a community or substructure. The augmentation step adds a $(K+1)$-th component $N(v;\mu_{K+1},\Sigma_{K+1})$; its placement is constrained by two inequalities. The novelty condition requires the KL divergence between the new component and every existing component to be at least $\delta_0$, and the reliability condition requires the KL divergence between the original GMM and the new GMM to be at most $\delta_1$, computed through a variational upper bound $\bar{d}_{\phi,\psi}$. Because the KL between two Gaussians has a closed form, $\mu_{K+1}$ and $\Sigma_{K+1}$ are updated by stochastic gradient ascent, with the covariance update carried out in the matrix-exponentiated (mirror descent) geometry. A decomposed normalized maximum likelihood check that the latent data with the new component are best described by $K+1$ clusters guards that the added blob is indeed a new cluster, and the decoder threshold is set to the average score of existing edges. This machinery turns "invent a new community" into a parameter search in the latent density.

What would settle it

Run GCA on a stochastic block model with known planted communities, then compute the modularity of the generated graph and the normalized mutual information between detected communities and the planted partition; alternatively, compare the edge density inside the added node set with the edge density between that set and the original nodes. If the added nodes do not form a block with higher internal than cross edge density, or if modularity does not increase relative to the original graph, the central claim that GCA generates a new community is falsified even when the latent mixture has a new Gaussian component.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that graph community augmentation can be carried out as a single well-chosen edit to the probability density that a graph generator has learned in latent space, rather than by retraining or by sampling far from the data. The paper claims that if a variational graph autoencoder embeds the nodes of a graph into a space where a Gaussian mixture model describes their positions, then adding a $(K+1)$-th Gaussian component whose parameters satisfy a novelty constraint (KL distance at least $\delta_0$ from every existing component) and a reliability constraint (KL distance at most $\delta_1$ from the original mixture), and decoding fresh samples from it, yields a graph with a new community that still inherits the basic nature of the original graph. In support, the authors report on synthetic stochastic block model graphs that GCA achieves higher valid-unique-novel percentages and much higher anomaly scores than GDSS, ConGress, and DiGress while staying comparable in degree, clustering, orbit, and spectral MMD, and on Cora, PubMed, CiteSeer, and Coauthor-Physics that it produces higher anomaly scores than GraphRNN and EDGE while remaining comparable in structural statistics.

Load-bearing premise

The load-bearing premise is that a Gaussian cluster in the latent space of a graph autoencoder corresponds to a graph community, so that decoding points sampled from a newly added Gaussian component creates a new community rather than merely a dense blob of extra nodes.

Editorial extensions

If this is right

  • A generator built this way can extrapolate to a new community at augmentation time without retraining: the same VGAE encoder and decoder serve both the original graph and the augmented one.
  • The two thresholds $\delta_0$ and $\delta_1$ expose a controllable novelty-versus-realism trade-off, so a user can choose where on that frontier to generate graphs.
  • The augmentation step is fast enough for large graphs: the reported complexity is linear in node and edge count and in cluster count, and augmentation on the real datasets takes under two minutes.
  • Because structure preservation is quantified by MMD on standard graph statistics, GCA can be compared, and is claimed comparable, to strong diffusion and recurrent generators on that axis.

Reading between the lines

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

  • Editorial inference: the paper does not measure community structure in the generated graphs, so its headline claim rests on an unverified mapping from latent Gaussian components to graph communities; computing modularity or normalized mutual information on synthetic graphs with planted communities would settle this.
  • Editorial inference: if that mapping fails, GCA would still generate structurally plausible novel graphs, but the "community" wording would overstate what is produced: the DNML check certifies a new Gaussian component, not necessarily a new community.
  • Editorial inference: the same density-edit recipe should transfer to other autoencoders, mixture families, or distance measures; the reusable idea is to treat extrapolation as adding a component to a latent mixture under novelty and reliability constraints.
  • Editorial inference: the inner-product decoder with an averaged threshold implies the generated community will appear as a dense, assortative subgraph; testing on disassortative or bipartite ground-truth communities would reveal whether the approach generalizes.
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.

Referee Report

5 major / 6 minor

Summary. The paper introduces the graph community augmentation problem and proposes an algorithm, GCA, that trains a VGAE on a graph, fits a GMM to the node embeddings, selects the number of latent clusters via a DNML-MDL criterion, and then adds a new Gaussian component. The new component's mean and covariance are optimized to satisfy a novelty condition (KL distance to existing clusters at least δ0) and a reliability condition (KL divergence between the original and augmented GMMs at most δ1, checked through a variational upper bound). The algorithm samples M latent points from the new component, concatenates them with the original embeddings, verifies in latent space that the augmented set has K+1 clusters, and decodes the result into a graph by thresholding cosine similarities. Experiments on an SBM dataset and four real graph datasets compare GCA with graph diffusion and autoregressive baselines using V.U.N., an Anomaly score, and MMD statistics of graph properties.

Significance. If the central community-level claim were directly validated, the paper would be a useful first step toward extrapolative graph generation, with a principled MDL-based cluster-number selection and an explicit formulation of novelty and reliability in latent space. The authors provide source code and evaluate on multiple datasets with several baselines. However, the current evidence does not establish that the added latent component decodes to a graph community, and the main novelty metric, Anomaly, is partly aligned with the optimization objective rather than an independent measure. The significance is therefore contingent on additional evaluation.

major comments (5)
  1. [Sec. V-A5, V-B5; Algorithm 2] The central claim that GCA generates 'a novel graph community inheriting the basic nature in the original graph' is not directly tested. Algorithm 2 (lines 6-9) verifies only that the concatenated latent set v||v' has K+1 Gaussian components under the DNML criterion, and Figures 3-4 show t-SNE separation in latent space. The decoded graph G' is never examined for community structure: no modularity, NMI, conductance, or internal/external edge-density comparison is reported for the M added nodes. A latent-space cluster count is not evidence that the decoded nodes form a community in the graph; they could form a disconnected component, merge with an existing community, or spread across communities. Please add direct community-structure metrics on the generated graphs, such as detected-community overlap with the added-node set, block-wise edge density, or NMI against SBM ground-truth blocks.
  2. [Sec. V-A3, Sec. IV-B2, Eq. (12)] The Anomaly metric is not an independent measure of novelty and is not comparable across methods. For GCA, Anomaly is defined as -log p_orig(v) in the latent space, while the augmentation step explicitly maximizes the KL distance from the existing GMM components (Eq. (12) and the SGA updates in Sec. IV-B2). The generated latent points are therefore placed in low-density regions of p_orig by construction, so a high Anomaly score is expected even if the decoded graph has no meaningful community. For the baselines, Anomaly is computed in different probability spaces (diffusion-model likelihoods of node features and adjacency), so the large Anomaly gaps in Tables I and II largely reflect different models rather than a validated property. Please use a common or external novelty measure, or include a control that samples from the fitted GMM without the augmentation step.
  3. [Algorithm 2, lines 6-9] The acceptance loop for Kest has no termination guarantee and can act as a favorable-random-draw selector. After the new component is optimized, the algorithm repeatedly samples M points and re-estimates Kest on the same concatenated latent set; if Kest is not K+1, it samples again without changing the optimization. Since the sample is random, the loop can in principle continue indefinitely until a draw happens to give K+1, and no maximum iterations or failure statistics are reported. Please add a stopping rule, report the number of iterations or failure rate, and, more importantly, verify the decoded graph's community structure rather than only the latent cluster count.
  4. [Table II] The claim of structural preservation is overstated for real datasets. On several MMD statistics GCA is substantially worse than EDGE, for example Cora Deg 0.0231 vs 0.0123, PubMed Orbit 0.2359 vs 0.1593, CiteSeer Orbit 0.2283 vs 0.1236, and Coauthor-Physics Spec 0.7058 vs 0.6365 for M=50. No error bars or significance tests are provided for the MMD values, so the statement that GCA is 'comparable to the rival algorithms' needs qualification and statistical support.
  5. [Table IV, Sec. V-D] The ablation study contains results that contradict the narrative in the text. In Table IV, replacing VGAE with GMMDA (row 3) and adding GMCD (rows 4 and 5) reduces Anomaly on Cora from 6.39 to 1.59, 0.66, and 1.55, respectively, whereas the text says GMMDA is 'more effective' and that GMCD causes only a 'slight' decrease. The SBM results show a different pattern. Please reconcile these discrepancies or explain why Cora behaves differently; the current discussion does not support the stated conclusions.
minor comments (6)
  1. [Eq. (11)] The loss function uses e_ij ∈ {0,1}, but the input adjacency matrix A is described as weighted; please clarify how weighted edges are binarized for training and decoding.
  2. [Algorithm 1] The Input line lists 'labels of nodes' but labels are not used anywhere in the algorithm; please remove them or explain their role.
  3. [Algorithm 2, line 10] The notation v⊕v' should be v∥v' to match Section III-A and Eq. (4).
  4. [Eq. (14), Appendix A] The quantity d_phi,psi is used in Eq. (14) before it is defined; please add a forward pointer to Appendix A or define it in the main text.
  5. [Sec. IV-B1, Eq. (18)] The use of the variational upper bound in Eq. (14) is conservative for the inequality in Eq. (18), since d_KL(p_orig,p_new) ≤ d_phi,psi, so the reliability check is not incorrect; however, reporting the estimated true KL divergence alongside the upper bound would help the reader interpret the tightness of δ1.
  6. [Sec. V-A4, V-B4] The hyperparameter setting states that initial values of the new cluster are set to those of an existing cluster, but it is unclear how the choice of which existing cluster affects results; please report sensitivity to this initialization or explain why it is not varied.

Circularity Check

1 steps flagged · score 6.0 of 10

Real-dataset novelty claim is circular: Anomaly is -log p_orig, the same density whose mixture components the augmentation step is optimized to move away from.

  1. fitted input called prediction [Sec. V-A3 (Anomaly metric); Sec. IV-B1-B2 (Eq. (12) and SGA update)]
    "we introduce an anomaly score (Anomaly) as the log-probability of the generated graphs for GCA: Anomaly= −logporig(v), where v∈RD is an embedding vector in a latent space and porig is the original probabilistic density function in Eq. (2). ... the novelty condition is expressed as follows for k=1,...,K : dKL(N (v;µk,Σk),N (v;µK+1,ΣK+1)) ≥ δ0, (12) ... We update µK+1 and ΣK+1 with the stochastic gradient ascent (SGA) ... This procedure is repeated until Eq. (12) is satisfied."

    The generated points v' are sampled only from N(µ_{K+1},Σ_{K+1}), whose parameters are optimized via SGA to satisfy Eq. (12), i.e. to be KL-distant from every existing component of porig. Anomaly is then defined as -log porig(v'), the negative log-likelihood of the same latent points under the original mixture. Moving the new Gaussian far from the existing components in KL places typical samples in low-density regions of porig, so high Anomaly is an expected artifact of the optimization, not independent evidence of novelty. On the real datasets (Table II), where V.U.N. is unavailable, this circular metric is the sole novelty measure used for the claim that GCA generated novel graphs. The synthetic V.U.N.

full rationale

The core circularity is confined to the novelty evaluation on real datasets. Anomaly is defined as -log porig(v), and the augmentation step explicitly optimizes µ_{K+1}, Σ_{K+1} so that the new Gaussian is far from every component of porig in KL divergence (Eq. (12)). A point sampled from a cluster that is forced to be KL-far from all existing clusters will generally fall in low-probability regions of porig, making Anomaly large. Therefore Tables II and IV's high Anomaly scores for GCA are a direct consequence of the algorithm's own constraint, not an independent confirmation that a genuinely novel graph community was generated. The synthetic experiments do provide external evidence via V.U.N. and MMD, so the overall circularity is partial rather than total. Separately, the paper's central claim of generating a 'new graph community' is under-supported because Algorithm 2 verifies only latent-space cluster count Kest == K+1 and never measures graph-level community structure (e.g., modularity, NMI, conductance); this is a validity gap, not a circularity, and is reflected here only as context. The DNML/VGAE machinery cited from prior work, including the authors' own papers, is not used as an unverified uniqueness theorem or to smuggle in the target conclusion, so self-citation is not load-bearing in this analysis.

Assumptions & free parameters 6 free parameters · 6 assumptions · 1 invented entities

The method rests on the assumptions that latent GMM clusters equal communities, that a new latent component decodes into a new community, and that the KL-based reliability bound preserves graph structure. These assumptions are plausible but unverified. The only tuned quantities are user thresholds and candidate ranges, not fitted parameters in the target metric, but several evaluation choices are ad hoc.

free parameters (6)
  • novelty threshold delta_0 = 5 (default)
    Specified by user; controls minimum KL distance from the new cluster to each existing cluster; sensitivity shown in Table III.
  • reliability threshold delta_1 = Dmax (max KL distance among existing clusters)
    Specified by user; bounds the KL shift between the original and augmented GMM; sensitivity shown in Table III.
  • M, number of added nodes = 5 to 25 (SBM), 50 to 200 (real datasets)
    Determines the size of the new community; no principled selection rule is given, and structural metrics degrade as M grows.
  • decoder edge threshold = average score of existing training edges
    Binarization threshold for the sigmoid output in Eq. (10), calibrated on training data.
  • VGAE learning rate and epochs = 0.01, 200
    Standard training hyperparameters; not central to the method but affect the learned embeddings.
  • candidate sets K and D = K={2,...,10}, D={4,8,16,24,32}
    User-provided ranges for model selection via DNML in Eq. (7).
assumptions (6)
  • domain assumption GMM components in the latent space correspond to graph communities
    Central to the claim; no community-level validation is provided. Invoked in Section IV-B and Figure 3.
  • domain assumption VGAE embeddings preserve community structure
    VGAE is designed for link prediction; the paper does not verify that embeddings align with communities. Used in Section IV-A1.
  • domain assumption The normalized inner-product decoder produces realistic graphs
    Eq. (10) is a simple inner-product link model with no structural guarantee; used to decode generated latent points into graphs.
  • domain assumption The variational KL upper bound is a sufficient proxy for the reliability condition
    Eq. (18) checks the upper bound d_{phi,psi}, not the true KL; if the bound is loose, the actual reliability may be violated. See Appendix A.
  • standard math DNML code-length correctly selects the number of clusters
    Cites [36], [37]; used in Eq. (7) and in the K_est check in Algorithm 2.
  • standard math Matrix exponentiated gradient and SGA converge to a valid cluster placement
    Cites [45], [46]; standard optimization tools, provided iterations terminate.
invented entities (1)
  • new latent-space Gaussian component
    purpose: Represents the new community to be added to the graph
    The component is a modeling construct; its existence is only checked by latent-space DNML (K_est == K+1) and t-SNE visualization, not by any external community measurement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Graph Community Augmentation with GMM-based Modeling in Latent Space." pith.science (2026). https://pith.science/paper/VPFROKJU

@misc{pith2026241201163,
  author       = {Pith},
  title        = {Pith review of: Graph Community Augmentation with GMM-based Modeling in Latent Space},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VPFROKJU}},
  note         = {Machine review of arXiv:2412.01163}
}
read the original abstract

This study addresses the issue of graph generation with generative models. In particular, we are concerned with graph community augmentation problem, which refers to the problem of generating unseen or unfamiliar graphs with a new community out of the probability distribution estimated with a given graph dataset. The graph community augmentation means that the generated graphs have a new community. There is a chance of discovering an unseen but important structure of graphs with a new community, for example, in a social network such as a purchaser network. Graph community augmentation may also be helpful for generalization of data mining models in a case where it is difficult to collect real graph data enough. In fact, there are many ways to generate a new community in an existing graph. It is desirable to discover a new graph with a new community beyond the given graph while we keep the structure of the original graphs to some extent for the generated graphs to be realistic. To this end, we propose an algorithm called the graph community augmentation (GCA). The key ideas of GCA are (i) to fit Gaussian mixture model (GMM) to data points in the latent space into which the nodes in the original graph are embedded, and (ii) to add data points in the new cluster in the latent space for generating a new community based on the minimum description length (MDL) principle. We empirically demonstrate the effectiveness of GCA for generating graphs with a new community structure on synthetic and real datasets.

Figures

Figures reproduced from arXiv: 2412.01163 by the authors.

Figure 1
Figure 1. Illustration of the graph community augmentation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall flow of the proposed algorithm called the graph [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. An example of the data points in the latent space and [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The data points in the latent space and the generated [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 49 canonical work pages

  1. [1]

    Deep gen- erative modelling: a comparative review of V AEs, GANs, normalizing flows, energy-based and autoregressive models,

    S. Bond-Taylor, A. Leach, Y . Long, and C. G. Willcocks, “Deep gen- erative modelling: a comparative review of V AEs, GANs, normalizing flows, energy-based and autoregressive models,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 44, pp. 7327–7347, 2022

    work page 2022

  2. [2]

    Graph generators: state of the art and open challenges,

    A. Bonifati, I. Holubov ´a, A. Prat-P ´erez, and S. Sakr, “Graph generators: state of the art and open challenges,” ACM Computing Surveys, vol. 53, no. 2, pp. 1–30, 2020

    work page 2020

  3. [3]

    Deep graph generators: a survey,

    F. Faez, Y . Ommi, M. S. Baghshah, and H. R. Rabiee, “Deep graph generators: a survey,” IEEE Access, vol. 9, pp. 106 675–106 702, 2021

    work page 2021

  4. [4]

    A systematic survey on deep generative models for graph generation,

    X. Guo and L. Zhao, “A systematic survey on deep generative models for graph generation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 5, pp. 5370–5390, 2022

    2022

  5. [5]

    A survey on deep graph generation: methods and applications,

    Y . Zhu, Y . Du, Y . Wang, Y . Xu, J. Zhang, Q. Liu, and S. Wu, “A survey on deep graph generation: methods and applications,” in Proc. of LoG , 2022

    work page 2022

  6. [6]

    Data augmentation for deep graph learning: a survey,

    K. Ding, Z. Xu, H. Tong, and H. Liu, “Data augmentation for deep graph learning: a survey,” ACM SIGKDD Explorations Newsletter , vol. 24, no. 2, pp. 61–77, 2022

    work page 2022

  7. [7]

    Graph data augmentation for graph machine learning: a survey,

    T. Zhao, W. Jin, Y . Liu, Y . Wang, G. Liu, S. G ¨unneman, N. Shah, and M. Jiang, “Graph data augmentation for graph machine learning: a survey,” IEEE Data Engineering Bulletin , vol. 46, no. 2, pp. 140–165, 2023

    work page 2023

  8. [8]

    Keygraph: automatic index- ing by co-occurrence graph based on building construction metaphor,

    Y . Ohsawa, N. E. Benson, and M. Yachida, “Keygraph: automatic index- ing by co-occurrence graph based on building construction metaphor,” in Proceedings IEEE International Forum on Research and Technology Advances in Digital Libraries (ADL) , 1998, pp. 12–18

    work page 1998

Show all 56 references
  1. [9]

    Chance discoveries for making decisions in complex real world,

    Y . Ohsawa, “Chance discoveries for making decisions in complex real world,” New Generation Computing , no. 20, pp. 143–163, 2002

    2002

  2. [10]

    J. K. Harris, An introduction to exponential random graph modeling . SAGE Publications, 2014

    2014

  3. [11]

    Learning community embedding with community detection and node embedding on graphs,

    S. Cavallari, V . W. Zheng, H. Cai, K. C.-C. Chang, and E. Cambria, “Learning community embedding with community detection and node embedding on graphs,” in Proc. of CIKM , 2017, pp. 377–386

    2017

  4. [12]

    Deep clustering by Gaussian mixture variational autoencoders with graph embedding,

    L. Yang, N.-M. Cheung, J. Li, and J. Fang, “Deep clustering by Gaussian mixture variational autoencoders with graph embedding,” in Proc. of ICCV, 2019, pp. 6440–6449

    2019

  5. [13]

    Principled knowledge extrapolation with GANs,

    R. Feng, J. Xiao, K. Zheng, D. Zhao, J. Zhou, Q. Sun, and Z.-J. Zha, “Principled knowledge extrapolation with GANs,” in Proc. of ICML , 2021, pp. 6447–6464

    2021

  6. [14]

    Deep extrapolation for attribute-enhanced generation,

    A. Chan, A. Madani, B. Krause, and N. Naik, “Deep extrapolation for attribute-enhanced generation,” in Proc. of NeurIPS , 2021, pp. 14 084– 14 096

    2021

  7. [15]

    A theory of independent mechanisms for extrapolation in generative models,

    M. Besserve, R. Sun, D. Janzing, and B. Scholkopf, “A theory of independent mechanisms for extrapolation in generative models,” in Proc. of AAAI , 2021, pp. 6741–6749

    2021

  8. [16]

    Controllable and progressive image extrapolation,

    Y . Li, L. Jiang, and M.-H. Yang, “Controllable and progressive image extrapolation,” in Proc. of WACV, 2021, pp. 2140–2149

    2021

  9. [17]

    A deep generative approach to search extrapolation and recommendation,

    F. X. Han, D. Niu, H. Chen, K. Lai, Y . He, and Y . Xu, “A deep generative approach to search extrapolation and recommendation,” inProc. of KDD, 2019, pp. 1771–1779

    2019

  10. [18]

    Half-Hop: a graph upsampling approach for slowing down message passing,

    M. Azabou, V . Ganesh, S. Thakoor, C.-H. Lin, L. Sathidevi, R. Liu, M. Valko, P. Veliˇckovi´c, and E. L. Dyer, “Half-Hop: a graph upsampling approach for slowing down message passing,” in Proc. of ICML , 2023, pp. 1341–1360

    2023

  11. [19]

    Local augmentation for graph neural networks,

    S. Liu, R. Ying, H. Dong, L. Li, T. Xu, Y . Rong, P. Zhao, J. Huang, and D. Wu, “Local augmentation for graph neural networks,” in Proc. of ICML, 2022, pp. 14 054–14 072

    2022

  12. [20]

    GMMDA: Gaussian mixture modeling of graph in latent space for graph data augmentation,

    Y . Li, L. Xu, and K. Yamanishi, “GMMDA: Gaussian mixture modeling of graph in latent space for graph data augmentation,” in Proc. of ICDM, 2023, pp. 319–328

    2023

  13. [21]

    Spectral augmentation for self-supervised learning on graphs,

    L. Lin, J. Chen, and H. Wang, “Spectral augmentation for self-supervised learning on graphs,” in Proc. of ICLR , 2022

    2022

  14. [22]

    Graph self-supervised learning with accurate discrepancy learning,

    D. Kim, J. Baek, and S. J. Hwang, “Graph self-supervised learning with accurate discrepancy learning,” in Proc. of NeurIPS , 2022

    2022

  15. [23]

    Out-of-distribution generaliza- tion on graphs: a survey,

    H. Li, X. Wang, Z. Zhang, and W. Zhu, “Out-of-distribution generaliza- tion on graphs: a survey,” arXiv preprint 2202.07987, 2022

    2022 arXiv

  16. [24]

    Causal- GAN: Learning causal implicit generative models with adversarial training,

    M. Kocaoglu, C. Snyder, A. G. Dimakis, and S. Vishwanath, “Causal- GAN: Learning causal implicit generative models with adversarial training,” in Proc. of ICLR , 2018

    2018

  17. [25]

    Counterfactual generative networks,

    A. Sauer and A. Geiger, “Counterfactual generative networks,” in Proc. of ICLR, 2020

    2020

  18. [26]

    CounteRGAN: Generating realistic counterfactuals with residual generative adversarial nets,

    D. Nemirovsky, N. Thiebaut, Y . Xu, and A. Gupta, “CounteRGAN: Generating realistic counterfactuals with residual generative adversarial nets,” arXiv preprint arXiv:2009.05199 , 2020

    2009 arXiv

  19. [27]

    CausalV AE: disentangled representation learning via neural structural causal models,

    M. Yang, F. Liu, Z. Chen, X. Shen, J. Hao, and J. Wang, “CausalV AE: disentangled representation learning via neural structural causal models,” in Proc. of CVPR , 2021, pp. 9593–9602

    2021

  20. [28]

    The counterfactual χ-GAN: finding comparable cohorts in observational health data,

    A. J. Averitt, N. Vanitchanant, R. Ranganath, and A. J. Perotte, “The counterfactual χ-GAN: finding comparable cohorts in observational health data,” Journal of Biomedical Informatics , vol. 109, p. 103515, 2020

    2020

  21. [29]

    Designing counterfactual generators using deep model inversion,

    J. J. Thiagarajan, V . Narayanaswamy, D. Rajan, J. Liang, A. Chaudhari, and A. Spanias, “Designing counterfactual generators using deep model inversion,” in Proc. of NeurIPS , vol. 34, 2021

    2021

  22. [30]

    Causal inference in statistics: An overview,

    J. Pearl, “Causal inference in statistics: An overview,” Statistics surveys, vol. 3, pp. 96–146, 2009

    2009

  23. [31]

    C. M. Bishop, Pattern Recognition and Machine Learning . Springer, 2006

    2006

  24. [32]

    Deep learning on graphs: a survey,

    Z. Zhang, P. Cui, and W. Zhu, “Deep learning on graphs: a survey,” IEEE Transactions on Knowledge and Data Engineering , vol. 34, pp. 249–270, 2022

    2022

  25. [33]

    Variational graph auto-encoders,

    T. N. Kipf and M. Welling, “Variational graph auto-encoders,” arXiv preprint 1611.07308, 2016

    2016 arXiv

  26. [34]

    Modeling by shortest data description,

    J. Rissanen, “Modeling by shortest data description,” Automatica, vol. 14, pp. 465–471, 1978

    1978

  27. [35]

    Yamanishi, Learning with the Minimum Description Length Princi- ple

    K. Yamanishi, Learning with the Minimum Description Length Princi- ple. Springer Nature, 2023

    2023

  28. [36]

    Decomposed normalized maximum likelihood codelength criterion for selecting hierarchical latent variable models,

    T. Wu, S. Sugawara, and K. Yamanishi, “Decomposed normalized maximum likelihood codelength criterion for selecting hierarchical latent variable models,” in Proc. of KDD , 2017, pp. 1165–1174

    2017

  29. [37]

    The decomposed normalized maximum likelihood code-length criterion for selecting hier- archical latent variable models,

    K. Yamanishi, T. Wu, S. Sugawara, and M. Okada, “The decomposed normalized maximum likelihood code-length criterion for selecting hier- archical latent variable models,” Data Mining and Knowledge Discovery, vol. 33, pp. 1017–1058, 2019

    2019

  30. [38]

    Graph summarization with latent variable probabilistic models,

    S. Fukushima, R. Kanai, and K. Yamanishi, “Graph summarization with latent variable probabilistic models,” in Proc. of ComplexNetworks, 2021, pp. 428–440

    2021

  31. [39]

    Detecting hierarchical changes in latent variable model,

    S. Fukushima and K. Yamanishi, “Detecting hierarchical changes in latent variable model,” in Proc. of ICDM , 2020, pp. 1128–1133

    2020

  32. [40]

    Balancing summarization and change detection in graph streams,

    ——, “Balancing summarization and change detection in graph streams,” in Proc. of ICDM , 2023, pp. 1025–1030

    2023

  33. [41]

    Rissanen, Optimal estimation of parameters

    J. Rissanen, Optimal estimation of parameters . Cambridge, 2012

    2012

  34. [42]

    A linear-time algorithm for computing the multinomial stochastic complexity,

    P. Kontkanen and P. Myllym ¨aki, “A linear-time algorithm for computing the multinomial stochastic complexity,” Information Processing Letters, vol. 103, no. 6, pp. 227–233, 2007

    2007

  35. [43]

    Approximating the Kullback Leibler divergence between Gaussian mixture models,

    J. R. Hershey and P. A. Olsen, “Approximating the Kullback Leibler divergence between Gaussian mixture models,” in Proc. of ICASSP , 2007, pp. 317–320

    2007

  36. [44]

    Lower and upper bounds for approximation of the Kullback-Leibler divergence between Gaussian mixture models,

    J.-L. Durrieu, J.-P. Thiran, and F. Kelly, “Lower and upper bounds for approximation of the Kullback-Leibler divergence between Gaussian mixture models,” in Proc. of ICASSP , 2012, pp. 4833–4836

    2012

  37. [45]

    Matrix exponentiated gradient updates for on-line learning and Bregman projection,

    K. Tsuda, G. Ratsch, and M. K. Warmuth, “Matrix exponentiated gradient updates for on-line learning and Bregman projection,” Journal of Machine Learning Research , vol. 6, pp. 995–1018, 2005

    2005

  38. [46]

    Exponentiated gradient versus gradient descent for linear predictors,

    J. Kivinen and M. K. Warmuth, “Exponentiated gradient versus gradient descent for linear predictors,” Information and Computation , vol. 132, no. 1, pp. 1–63, 1997

    1997

  39. [47]

    SPECTRE: spectral conditioning helps to overcome the expressivity limits of one- shot graph generators,

    K. Martinkus, A. Loukas, N. Perraudin, and R. Wattenhofer, “SPECTRE: spectral conditioning helps to overcome the expressivity limits of one- shot graph generators,” in Proc. of ICML , 2022, pp. 15 159–15 179

    2022

  40. [48]

    Graph generation with diffusion mixture,

    J. Jo, D. Kim, and S. J. Hwang, “Graph generation with diffusion mixture,” in Proc. of ICML , 2024, pp. 22 371–22 405

    2024

  41. [49]

    Estimation and prediction for stochastic blockmodels for graphs with latent block structure,

    T. A. B. Snijders, “Estimation and prediction for stochastic blockmodels for graphs with latent block structure,” Journal of Classification, vol. 14, pp. 75–100, 1997

    1997

  42. [50]

    Score-based generative modeling of graphs via the system of stochastic differential equations,

    J. Jo, S. Lee, and S. J. Hwang, “Score-based generative modeling of graphs via the system of stochastic differential equations,” in Proc. of ICML, 2022, pp. 10 362–10 383

    2022

  43. [51]

    Digress: discrete denoising diffusion for graph generation,

    C. Vignac, I. Krawczuk, A. Siraudin, B. Wang, V . Cevher, and P. Frossard, “Digress: discrete denoising diffusion for graph generation,” in Proc. of ICLR , 2023

    2023

  44. [52]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Proc. of NeurIPS , 2020, pp. 6840–6851

    2020

  45. [53]

    Visualizing data using t-sne,

    L. van der Maaten and G. Hinton, “Visualizing data using t-sne,” Journal of Machine Learning Research , vol. 9, pp. 2579–2605, 2008

    2008

  46. [54]

    Graphrnn: generating realistic graphs with deep auto-regressive models,

    J. You, R. Ying, X. Ren, W. L. Hamilton, and J. Leskovec, “Graphrnn: generating realistic graphs with deep auto-regressive models,” in Proc. of ICML, 2018, pp. 5708–5717

    2018

  47. [55]

    Efficient and degree-guided graph generation via discrete diffusion modeling,

    H. Chen, J. He, X. Han, and L.-P. Liu, “Efficient and degree-guided graph generation via discrete diffusion modeling,” in Proc. of ICML , 2023, pp. 4585–4610

    2023

  48. [56]

    Diffusing gaussian mixtures for generating categorical data,

    F. Regol and M. Coates, “Diffusing gaussian mixtures for generating categorical data,” in Proc. of AAAI , 2023, pp. 9570–9578. APPENDIX A. Upper Bound of KL Divergence between GMMs KL divergence between two GMMs in Eq. (14) is upper- bounded, in reference to [43], [44], as fol...

    2023

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.