SBGD: Improving Graph Diffusion Generative Model via Stochastic Block Diffusion

Junwei Su; Shan Wu

REVIEW 4 major objections 5 minor 60 references

SBGD: Improving Graph Diffusion Generative Model via Stochastic Block Diffusion

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Running graph diffusion on block subgraphs rather than the full graph cuts memory complexity from O(N^2) to O(C^2 + C F) while keeping generation quality and improving size extrapolation.

desk verdict A useful block-decomposition idea undercut by a mismatch between the stated diffusion factorization and the implemented algorithm; worth a major revision, not acceptance as-is. read the letter →

arxiv 2508.14352 v1 pith:PSXG3FIP submitted 2025-08-20 cs.LG

Junwei Su , Shan Wu This is my paper

classification cs.LG

keywords graphgenerationdiffusiongenerativemodelsstochasticblockmodeldecompositionmemoryefficiencysizegeneralizationtransformers

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

SBGD is a graph diffusion model that runs the forward and reverse diffusion process not on the whole graph but on a decomposition into block subgraphs, with a separate lightweight network that predicts edges between blocks. The paper's claim is that this block space preserves generation quality while cutting memory complexity from quadratic in the number of nodes to quadratic in the block size, enabling training on graphs that make full-graph methods run out of memory. Empirically, SBGD reports up to a 6x memory-ratio improvement over state-of-the-art baselines with comparable or better generative metrics, and its block representation improves extrapolation to graph sizes not seen in training. A sympathetic reader would care because the obstacle that stops graph diffusion models from scaling to large real-world graphs is exactly the quadratic memory footprint that block decomposition removes, and the size-generalization gain suggests the model learns transferable local structure rather than memorizing a fixed graph size.

What carries the argument

The load-bearing object is the block graph representation: the graph is partitioned into k small subgraphs (blocks) that are internally dense and mutually sparse, following the stochastic-block-model prior of real networks. Diffusion is applied block-wise, and the inter-block adjacency is generated by a dedicated network that takes two generated blocks as input and predicts the edges between them. The block space is what converts O(N^2) memory to O(C^2 + C F), because no model component ever sees the full graph at once; it is also what makes size extrapolation plausible, because blocks are size-invariant building blocks that can be sampled in any number and recomposed into graphs of differen

What would settle it

Train SBGD on a synthetic block-structured graph family with a deliberately inserted global constraint invisible to any pair of blocks (for example, all inter-block edges are constrained to form a single Hamiltonian cycle, or the graph must satisfy a prescribed cross-block triangle count). If the generated graphs fail to reproduce that global property while a full-graph diffusion model succeeds, the independent-block factorization is the cause.

Watch

Extended reading notes

Core claim

The central discovery is that a graph diffusion generator need not model the full adjacency matrix as one monolithic object. SBGD partitions nodes into k blocks, diffuses each block's adjacency and feature matrices independently with Gaussian noise, and models inter-block edges with a separate network that takes two generated blocks as input and outputs their sparse connection pattern. The forward process factorizes as P(G(t)|G(t-1)) = product over blocks of the block transition distribution times product over block pairs of the inter-block transition distribution. Because each block is far smaller than the whole graph, memory scales as O(C^2 + C F) instead of O(N^2); because the blocks them

Load-bearing premise

The load-bearing premise is that a real graph's structure can be captured by independent block-local diffusion plus pairwise inter-block edges, so no important pattern requires coordinated structure across three or more blocks at once.

Editorial extensions

If this is right

Training and sampling can be run on graphs that full-graph GDGMs cannot fit in memory; the paper reports being the only method able to train on OGBN-products.
Memory savings grow with graph size, since the full-graph term is quadratic in N while the block term is quadratic in the chosen block size.
Generation can extrapolate to larger (and smaller) sizes than seen in training, because the same learned blocks can be reused in different numbers.
The partition count k becomes a practical tuning knob: too few partitions keep undesired global noise, too many destroy global structure, so an optimal granularity exists per dataset.
Distributed training becomes natural, because block pairs can be processed independently and in parallel across devices.

Reading between the lines

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

If the factorization is taken literally, any structure that requires coordinated edges among three or more blocks—long-range cycles, diameter constraints, global assortativity—will be under-modeled; a testable fix is to add a hierarchy of blocks or a global consistency term.
The paper's own ablation suggests block size should track the natural community scale of the data; one could replace the fixed partition count with an adaptive partitioner that chooses resolution from a spectral or modularity signal.
The modularization principle is generic: the same 'diffuse locally, predict interactions separately' recipe could apply to point clouds, program ASTs, or any structured object that decomposes into semi-independent parts.
The reported 6x memory ratio is measured on current hardware and datasets; because the asymptotic gap widens with N, on much larger graphs the practical savings should exceed 6x, provided the block structure remains valid.

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.

Desk Editor's Note

read the letter

The core idea is worth taking seriously: decompose the graph into stochastic blocks, diffuse each block, and predict inter-block edges with a lightweight network. That is a natural way to attack the O(N^2) memory bottleneck of graph diffusion models, and the results on OGBN-products (where baselines OOM) suggest the approach has real legs. The paper is also clearly written and the experimental setup is fairly standard, with multiple datasets and metrics. Credit where due: the memory complexity table is a useful contribution, and the authors make a genuine attempt to connect an old structural prior to a modern generative framework. The soft spots are substantial, though. The most load-bearing problem is that Eq. 3.1 factorizes the forward process as a product that includes diffusions on inter-block matrices A_ij, but Algorithm 1 never noises A_ij and Algorithm 2 obtains A_ij with a single forward pass of s_phi. So the claimed diffusion on inter-block connections is not what is implemented. That is not a minor detail — the memory savings and the size-generalization story both depend on generating many blocks and recombining them, and the paper never defines that recombination. Algorithm 2 samples exactly two blocks and stops; there is no multi-block assembly procedure. The size-generalization experiments also rely on cSBM, a dataset whose generative model matches the method’s block prior, so those results are contaminated by selection bias, with no error bars on the FID curves. On top of that, the abstract says “up to 6x” memory improvement, but Table 2 lists an 85.3x ratio for EDGE; either way it is an internal inconsistency. No code or data are provided, which makes it hard to verify anything beyond the stated numbers. All of this said, the idea is promising and the flaws are fixable in principle. A serious referee could help the authors tighten the method: either implement the inter-block diffusion properly or drop the claim that Eq. 3.1 is realized, specify the multi-block sampling, and rerun the size generalization on datasets that are not themselves generated by an SBM. I would not accept the paper as submitted, but I would send it to review rather than desk-reject it, because the core direction deserves engagement.

Referee Report

4 major / 5 minor

Summary. The paper proposes SBGD, a stochastic block graph diffusion model that decomposes a graph into k block graphs and pairwise inter-block adjacency matrices, then runs diffusion on the block graphs and predicts inter-block edges with a separate network. The central claims are (i) reduced memory complexity from O(N^2) to O(C^2 + CF), (ii) comparable or better generation quality, and (iii) improved size generalization through recombining learned blocks. Experiments on planar graphs, cSBM, QM9, OGBN-Arxiv, and OGBN-Products report up to a 6x memory ratio improvement, competitive FID/MMD scores, and better extrapolation to unseen graph sizes. The idea of modularizing graph diffusion via a stochastic block prior is interesting. However, as written, the training and sampling algorithms do not implement the forward process defined in Eq. (3.1): inter-block matrices are never diffused, the sampler produces only two blocks, and no procedure is given for assembling a full k-block graph. The cSBM size-generalization experiment also evaluates the model on data that matches its own structural prior. These issues directly undermine the paper's main claims.

Significance. If the proposed block-diffusion framework were fully realized, it could be a useful step toward scalable graph diffusion and size generalization: the memory complexity reduction from full-graph diffusion to per-block diffusion is a natural and potentially practical idea, and the empirical claim of being the only method to train on OGBN-Products is notable. The paper also makes its structural prior explicit through Eq. (3.1) and provides a memory-complexity comparison table. I credit the authors for identifying a real limitation of GDGMs and for attempting a modular decomposition. However, the significance is currently not established because the implementation and evaluation do not support the claimed mechanism: inter-block edges are generated one-shot, not by diffusion, and the multi-block assembly required for scalability and size generalization is absent.

major comments (4)

[Sec. 3.2, Eq. (3.1) vs. Algorithm 1] The forward process in Eq. (3.1) explicitly includes P(A_ij(t)|A_ij(t-1)) for every inter-block pair, meaning inter-block edges should undergo a diffusion process with a time index. Algorithm 1 never corrupts A_ij: it samples noise for A_i, A_j, X_i, X_j only, and the inter-block objective L_I is E[||s_phi(C_i,C_j)-A_ij||^2], a one-shot regression from clean predicted blocks. No A_ij(t) is fed to s_phi and no t conditioning is used. Therefore the trained model does not reverse a diffusion process on inter-block connections. This is load-bearing because the paper's central claim is that diffusion is performed in the block graph space including inter-block interactions; as implemented, only intra-block matrices are diffused.
[Algorithm 2 and Sec. 3.3] The sampling algorithm generates exactly two blocks, C_i and C_j, and a single inter-block matrix A_ij. There is no loop over k blocks, no procedure for choosing block sizes for k>2, and no rule for assembling the full graph from k blocks and k(k-1)/2 inter-block matrices. The memory complexity claim O(C^2+CF) in Table 1 and the size-generalization claim both rely on decomposing a large graph into many blocks and recombining them, but the paper never defines this recombination. For a graph of N=kC nodes, the full adjacency has k^2 C^2 entries; the O(C^2) figure ignores the number of blocks and the storage required for the final graph. As written, the scalability result is not demonstrated beyond two small blocks.
[Sec. 4.2, size generalization experiments] The size-generalization experiments are conducted on cSBM, a synthetic dataset generated from a stochastic block model with planted community structure. SBGD's method is exactly a block/community prior: it partitions nodes into blocks and models intra-block and inter-block connectivity. Evaluating on cSBM therefore measures how well the model recovers the same prior used to construct both the data and the model; it cannot distinguish learned extrapolation from the built-in structural assumption. A fair test would use graphs whose generative process is not block-structured, or at least compare against baselines on a non-SBM size-generalization benchmark. Without that, the claim that SBGD 'extrapolates better in size generation' is not supported.
[Sec. 3.3] There is an internal inconsistency in the memory analysis. Section 3.3 states the memory complexity is O(C^2), while Table 1 lists O(C^2 + CF). More importantly, neither expression includes the number of blocks k or the cost of storing and assembling the k(k-1)/2 inter-block matrices. If the method truly needs only O(C^2+CF) memory for the whole graph, this needs a precise accounting of how blocks and inter-block matrices are stored and processed; if it needs O(k^2 C^2) to store the final graph, then the claimed advantage over O(N^2) is not demonstrated. This is a load-bearing point for the scalability claim and should be clarified or corrected.

minor comments (5)

[Sec. 4.2] The text refers to 'Table 1' for the performance comparison, but the actual performance table is Table 2. The memory-complexity table is Table 1. Please fix the cross-reference.
[Figure 1] The caption lists Fig. 1(b) twice: 'Fig. 1(b) is a visualization of the example graph with...' and then 'Fig. 1(b) is a visualization of the adjacent matrix...' The second should be Fig. 1(c).
[Appendix C.2] The sentence 'Further exploration of partition algorithm selection and its impact on the performance of SGBD' contains a typo: 'SGBD' should be 'SBGD'.
[Sec. 3.3] The phrase 'the benefit of using analogue bit' is mentioned but not defined or connected to the rest of the section. Either elaborate or remove.
[General] The implementation details for the graph transformer backbone, the partition algorithm (METIS), and hyperparameters are deferred to the appendix, but the appendix says the 'detailed and complete procedure' is in the supplementary material, which is not included in the arXiv version. Please provide the full pseudo-code and open-source code to make the experiments reproducible.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central claims are empirical, and the self-citations are background motivation rather than load-bearing derivation.

full rationale

SBGD's central claims are empirical. The training objective (L = LAi + LAj + LXi + LXj + LI, Sec. 3.2) and sampling procedure (Algorithms 1-2) use standard denoising losses; no numerical result is obtained by plugging a fitted constant back into the same equation. The O(C^2+CF) memory claim is a complexity accounting for the block representation chosen in Sec. 3.1, not a prediction derived from data. Size generalization is demonstrated by FID experiments (Sec. 4.2), and the 'theoretical discussion' in Sec. 3.3 is explicitly a motivation, not a derivation. The self-citations to Su & Marbach (2022, 2023) appear only in background claims about real graphs having block structure, alongside many independent references (Abbe 2018; Newman 2006; Karrer & Newman 2011); they are not load-bearing. The cSBM dataset does share the method's SBM-like prior, which is an evaluation confound, but the paper does not fit anything to cSBM and then call it prediction. There are serious non-circular implementation gaps: Eq. 3.1 defines a diffusion for inter-block A_ij, but Algorithm 1 never corrupts A_ij and Algorithm 2 obtains bA_ij by a single forward pass of s_phi; moreover Algorithm 2 samples only two blocks, so the claimed multi-block recombination and O(C^2) memory for full-graph generation are not realized. These are correctness/completeness problems, not derivation circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method rests on a strong structural prior: graphs are assumed to be composed of dense blocks with sparse connections. The factorization into block-local diffusions plus pairwise inter-block predictors is the key ad hoc modeling choice. METIS partitioning and continuous Gaussian diffusion on discrete adjacency matrices are imported from prior work. The central hyperparameter k is tuned per dataset, and the edge-binarization step is unspecified.

free parameters (2)

Number of partitions k (block size C) = Not reported per dataset; tuned by grid search
The central hyperparameter controls the memory-quality tradeoff and the size generalization behavior (Fig. 4(b), 4(c)); the final values used in the reported results are not disclosed.
Edge binarization threshold for inter-block predictions = Unknown
s_phi outputs continuous entries for A_ij; the conversion to discrete edges (e.g., thresholding or sampling) is never specified, so any threshold is an implicit free parameter.

assumptions (4)

domain assumption Real-world graphs have block structure with dense intra-block and sparse inter-block connections.
Invoked throughout (Sec. 2.2, Sec. 3.3) to justify block decomposition and the light-weight inter-block module.
ad hoc to paper The graph distribution factorizes into a product of block-local distributions and pairwise inter-block interactions (Eq. 3.1).
This conditional independence-like factorization is the design premise of SBGD; if false, generating blocks independently and predicting edges pairwise loses global structure.
domain assumption METIS partitions capture the generative block structure such that block graphs from different graphs and sizes are exchangeable.
The method relies on METIS (Appendix C) to produce blocks whose distribution is stable across graph sizes for size generalization.
domain assumption Gaussian continuous diffusion on binary adjacency matrices is a valid approximation.
Adopted from GDSS/Jo et al. 2022; the paper applies Gaussian noise to A and X (Sec. 3.2), ignoring discreteness.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SBGD: Improving Graph Diffusion Generative Model via Stochastic Block Diffusion." pith.science (2026). https://pith.science/paper/PSXG3FIP

@misc{pith2026250814352,
  author       = {Pith},
  title        = {Pith review of: SBGD: Improving Graph Diffusion Generative Model via Stochastic Block Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PSXG3FIP}},
  note         = {Machine review of arXiv:2508.14352}
}

abstract

Graph diffusion generative models (GDGMs) have emerged as powerful tools for generating high-quality graphs. However, their broader adoption faces challenges in \emph{scalability and size generalization}. GDGMs struggle to scale to large graphs due to their high memory requirements, as they typically operate in the full graph space, requiring the entire graph to be stored in memory during training and inference. This constraint limits their feasibility for large-scale real-world graphs. GDGMs also exhibit poor size generalization, with limited ability to generate graphs of sizes different from those in the training data, restricting their adaptability across diverse applications. To address these challenges, we propose the stochastic block graph diffusion (SBGD) model, which refines graph representations into a block graph space. This space incorporates structural priors based on real-world graph patterns, significantly reducing memory complexity and enabling scalability to large graphs. The block representation also improves size generalization by capturing fundamental graph structures. Empirical results show that SBGD achieves significant memory improvements (up to 6$\times$) while maintaining comparable or even superior graph generation performance relative to state-of-the-art methods. Furthermore, experiments demonstrate that SBGD better generalizes to unseen graph sizes. The significance of SBGD extends beyond being a scalable and effective GDGM; it also exemplifies the principle of modularization in generative modeling, offering a new avenue for exploring generative models by decomposing complex tasks into more manageable components.

Figures

Figures reproduced from arXiv: 2508.14352 by the authors.

**Figure 2.** The overview of SBGD. (1) incoming graphs are decomposed into block graphs where vertices share similar properties. (2) sample from the set of block graphs and conduct separate diffusion. The noise model is defined by the distribution P(.) (3) The denoising network sθ, sψ, sϕ learns to predict the clean graph, including the inter-connections, from C (t) j , and C (t) i . During inference, the predicted distribution … view at source ↗

**Figure 3.** Visualization of graphs of different sizes generated from SBGD. The figure illustrates that SBGD is able to maintain the overall characteristic of the ground-truth (training graph) nicely even if generating graph of varied size. 0 10 20 Graph Size 0 30 60 FID Score (Lower is Better) GraphRNN SPECTRE SBGDD GDSS DiGress (a) Size Generalization of Different Methods 2 4 6 8 10 Number of Partitions 0 10 20 30 FID Score (… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Experiments on Size Generalization and Partition Number [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 40 canonical work pages

[1]

Community detection and stochastic block models: recent developments

Abbe, E. Community detection and stochastic block models: recent developments. Journal of Machine Learning Research, 18 0 (177): 0 1--86, 2018

work page 2018
[2]

and Mostajabdaveh, M

Aref, S. and Mostajabdaveh, M. Analyzing modularity maximization in approximation, heuristic, and graph neural network algorithms for community detection. Journal of Computational Science, 78: 0 102283, 2024

work page 2024
[3]

D., Ho, J., Tarlow, D., and Van Den Berg, R

Austin, J., Johnson, D. D., Ho, J., Tarlow, D., and Van Den Berg, R. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34: 0 17981--17993, 2021

work page 2021
[4]

and Albert, R

Barab \'a si, A.-L. and Albert, R. Emergence of scaling in random networks. science, 286 0 (5439): 0 509--512, 1999

work page 1999
[5]

L., and Polozov, O

Brockschmidt, M., Allamanis, M., Gaunt, A. L., and Polozov, O. Generative code modeling with graphs. arXiv preprint arXiv:1805.08490, 2018

arXiv 2018
[6]

Analog bits: Generating discrete data using diffusion models with self-conditioning

Chen, T., Zhang, R., and Hinton, G. Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202, 2022

arXiv 2022
[7]

Efficient and degree-guided graph generation via discrete diffusion modeling

Chen, X., He, J., Han, X., and Liu, L.-P. Efficient and degree-guided graph generation via discrete diffusion modeling. arXiv preprint arXiv:2305.04111, 2023

arXiv 2023
[8]

K., and Lu, X

Cherifi, H., Palla, G., Szymanski, B. K., and Lu, X. On community structure in complex networks: challenges and opportunities. Applied Network Science, 4 0 (1): 0 1--35, 2019

work page 2019

Show all 60 references

[9]

Cluster-gcn: An efficient algorithm for training deep and large graph convolutional networks

Chiang, W.-L., Liu, X., Si, S., Li, Y., Bengio, S., and Hsieh, C.-J. Cluster-gcn: An efficient algorithm for training deep and large graph convolutional networks. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pp.\ 257--266, 2019

2019
[10]

Chung, F. R. Spectral graph theory, volume 92. American Mathematical Soc., 1997

1997
[11]

Contextual stochastic block models

Deshpande, Y., Sen, S., Montanari, A., and Mossel, E. Contextual stochastic block models. Advances in Neural Information Processing Systems, 31, 2018

2018
[12]

and Nichol, A

Dhariwal, P. and Nichol, A. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34: 0 8780--8794, 2021

2021
[13]

Dobson, P. D. and Doig, A. J. Distinguishing enzyme structures from non-enzymes without alignments. Journal of molecular biology, 330 0 (4): 0 771--783, 2003

2003
[14]

On the evolution of random graphs

Erd o s, P., R \'e nyi, A., et al. On the evolution of random graphs. Publ. math. inst. hung. acad. sci, 5 0 (1): 0 17--60, 1960

1960
[15]

Algebraic connectivity of graphs

Fiedler, M. Algebraic connectivity of graphs. Czechoslovak mathematical journal, 23 0 (2): 0 298--305, 1973

1973
[16]

M., Rasch, M

Gretton, A., Borgwardt, K. M., Rasch, M. J., Sch \"o lkopf, B., and Smola, A. A kernel two-sample test. The Journal of Machine Learning Research, 13 0 (1): 0 723--773, 2012

2012
[17]

Graphite: Iterative generative modeling of graphs

Grover, A., Zweig, A., and Ermon, S. Graphite: Iterative generative modeling of graphs. In International conference on machine learning, pp.\ 2434--2444. PMLR, 2019

2019
[18]

L., Ying, R., and Leskovec, J

Hamilton, W. L., Ying, R., and Leskovec, J. Inductive representation learning on large graphs, 2018

2018
[19]

Gans trained by a two time-scale update rule converge to a local nash equilibrium

Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and Hochreiter, S. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017

2017
[20]

Denoising diffusion probabilistic models

Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

2020
[21]

W., Laskey, K

Holland, P. W., Laskey, K. B., and Leinhardt, S. Stochastic blockmodels: First steps. Social networks, 5 0 (2): 0 109--137, 1983

1983
[22]

Argmax flows and multinomial diffusion: Learning categorical distributions

Hoogeboom, E., Nielsen, D., Jaini, P., Forr \'e , P., and Welling, M. Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in Neural Information Processing Systems, 34: 0 12454--12465, 2021

2021
[23]

Open graph benchmark: Datasets for machine learning on graphs, 2021

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, 2021. URL https://arxiv.org/abs/2005.00687

2021 arXiv
[24]

Jo, J., Lee, S., and Hwang, S. J. Score-based generative modeling of graphs via the system of stochastic differential equations. In International Conference on Machine Learning, pp.\ 10362--10383. PMLR, 2022

2022
[25]

Jo, J., Kim, D., and Hwang, S. J. Graph generation with diffusion mixture. arXiv preprint arXiv:2302.03596, 2023

2023 arXiv
[26]

An algorithm for drawing general undirected graphs

Kamada, T., Kawai, S., et al. An algorithm for drawing general undirected graphs. Information processing letters, 31 0 (1): 0 7--15, 1989

1989
[27]

and Newman, M

Karrer, B. and Newman, M. E. Stochastic blockmodels and community structure in networks. Physical review E, 83 0 (1): 0 016107, 2011

2011
[28]

and Kumar, V

Karypis, G. and Kumar, V. Multilevelk-way partitioning scheme for irregular graphs. Journal of Parallel and Distributed computing, 48 0 (1): 0 96--129, 1998

1998
[29]

H., Vygen, J., Korte, B., and Vygen, J

Korte, B. H., Vygen, J., Korte, B., and Vygen, J. Combinatorial optimization, volume 1. Springer, 2011

2011
[30]

Lee, H., Hyung, E., and Hwang, S. J. Rapid neural architecture search by learning to generate graphs from datasets. arXiv preprint arXiv:2107.00860, 2021

2021 arXiv
[31]

Multi-objective de novo drug design with conditional graph generative model

Li, Y., Zhang, L., and Liu, Z. Multi-objective de novo drug design with conditional graph generative model. Journal of cheminformatics, 10: 0 1--24, 2018

2018
[32]

Spectre: Spectral conditioning helps to overcome the expressivity limits of one-shot graph generators

Martinkus, K., Loukas, A., Perraudin, N., and Wattenhofer, R. Spectre: Spectral conditioning helps to overcome the expressivity limits of one-shot graph generators. In International Conference on Machine Learning, pp.\ 15159--15179. PMLR, 2022

2022
[33]

Newman, M. E. Modularity and community structure in networks. Proceedings of the national academy of sciences, 103 0 (23): 0 8577--8582, 2006

2006
[34]

Newman, M. E. and Girvan, M. Finding and evaluating community structure in networks. Physical review E, 69 0 (2): 0 026113, 2004

2004
[35]

E., Watts, D

Newman, M. E., Watts, D. J., and Strogatz, S. H. Random graph models of social networks. Proceedings of the national academy of sciences, 99 0 (suppl\_1): 0 2566--2572, 2002

2002
[36]

Glide: Towards photorealistic image generation and editing with text-guided diffusion models

Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., and Chen, M. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021

2021 arXiv
[37]

Permutation invariant graph generation via score-based generative modeling

Niu, C., Song, Y., Song, J., Zhao, S., Grover, A., and Ermon, S. Permutation invariant graph generation via score-based generative modeling. In International Conference on Artificial Intelligence and Statistics, pp.\ 4474--4484. PMLR, 2020

2020
[38]

Deepwalk: Online learning of social representations

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

2014
[39]

Fr \'e chet chemnet distance: a metric for generative models for molecules in drug discovery

Preuer, K., Renz, P., Unterthiner, T., Hochreiter, S., and Klambauer, G. Fr \'e chet chemnet distance: a metric for generative models for molecules in drug discovery. Journal of chemical information and modeling, 58 0 (9): 0 1736--1741, 2018

2018
[40]

High-resolution image synthesis with latent diffusion models

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10684--10695, 2022

2022
[41]

and Komodakis, N

Simonovsky, M. and Komodakis, N. Graphvae: Towards generation of small graphs using variational autoencoders. In Artificial Neural Networks and Machine Learning--ICANN 2018: 27th International Conference on Artificial Neural Networks, Rhodes, Greece, October 4-7, 2018, Proceed...

2018
[42]

Deep unsupervised learning using nonequilibrium thermodynamics

Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pp.\ 2256--2265. PMLR, 2015

2015
[43]

Denoising diffusion implicit models

Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020 a

2010 arXiv
[44]

and Ermon, S

Song, Y. and Ermon, S. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019

2019
[45]

and Ermon, S

Song, Y. and Ermon, S. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33: 0 12438--12448, 2020

2020
[46]

P., Kumar, A., Ermon, S., and Poole, B

Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020 b

2011 arXiv
[47]

and Marbach, P

Su, J. and Marbach, P. Structure of core-periphery communities. In International Conference on Complex Networks and Their Applications, pp.\ 151--161. Springer, 2022

2022
[48]

and Marbach, P

Su, J. and Marbach, P. Structural properties of core--periphery communities. Advances in Complex Systems, pp.\ 2340004, 2023

2023
[49]

Thompson, R., Knyazev, B., Ghalebi, E., Kim, J., and Taylor, G. W. On evaluation metrics for graph generative models. arXiv preprint arXiv:2201.09871, 2022

2022 arXiv
[50]

Digress: Discrete denoising diffusion for graph generation

Vignac, C., Krawczuk, I., Siraudin, A., Wang, B., Cevher, V., and Frossard, P. Digress: Discrete denoising diffusion for graph generation. arXiv preprint arXiv:2209.14734, 2022

2022 arXiv
[51]

Graphgan: Graph representation learning with generative adversarial nets

Wang, H., Wang, J., Wang, J., Zhao, M., Zhang, W., Zhang, F., Xie, X., and Guo, M. Graphgan: Graph representation learning with generative adversarial nets. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018

2018
[52]

Deep graph library: A graph-centric, highly-performant package for graph neural networks

Wang, M., Zheng, D., Ye, Z., Gan, Q., Li, M., Song, X., Zhou, J., Ma, C., Yu, L., Gai, Y., et al. Deep graph library: A graph-centric, highly-performant package for graph neural networks. arXiv preprint arXiv:1909.01315, 2019

1909 arXiv
[53]

A compact review of molecular property prediction with graph neural networks

Wieder, O., Kohlbacher, S., Kuenemann, M., Garon, A., Ducrot, P., Seidel, T., and Langer, T. A compact review of molecular property prediction with graph neural networks. Drug Discovery Today: Technologies, 37: 0 1--12, 2020

2020
[54]

N., Gomes, J., Geniesse, C., Pappu, A

Wu, Z., Ramsundar, B., Feinberg, E. N., Gomes, J., Geniesse, C., Pappu, A. S., Leswing, K., and Pande, V. Moleculenet: a benchmark for molecular machine learning. Chemical science, 9 0 (2): 0 513--530, 2018

2018
[55]

Exploring randomly wired neural networks for image recognition

Xie, S., Kirillov, A., Girshick, R., and He, K. Exploring randomly wired neural networks for image recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 1284--1293, 2019

2019
[56]

Poisson flow generative models, 2022

Xu, Y., Liu, Z., Tegmark, M., and Jaakkola, T. Poisson flow generative models, 2022

2022
[57]

Diffsound: Discrete diffusion model for text-to-sound generation

Yang, D., Yu, J., Wang, H., Wang, W., Weng, C., Zou, Y., and Yu, D. Diffsound: Discrete diffusion model for text-to-sound generation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2023

2023
[58]

Graphrnn: Generating realistic graphs with deep auto-regressive models

You, J., Ying, R., Ren, X., Hamilton, W., and Leskovec, J. Graphrnn: Generating realistic graphs with deep auto-regressive models. In International conference on machine learning, pp.\ 5708--5717. PMLR, 2018

2018
[59]

Pard: Permutation-invariant autoregressive diffusion for graph generation

Zhao, L., Ding, X., and Akoglu, L. Pard: Permutation-invariant autoregressive diffusion for graph generation. arXiv preprint arXiv:2402.03687, 2024

2024 arXiv
[60]

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 gl...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Community detection and stochastic block models: recent developments

Abbe, E. Community detection and stochastic block models: recent developments. Journal of Machine Learning Research, 18 0 (177): 0 1--86, 2018

work page 2018

[2] [2]

and Mostajabdaveh, M

Aref, S. and Mostajabdaveh, M. Analyzing modularity maximization in approximation, heuristic, and graph neural network algorithms for community detection. Journal of Computational Science, 78: 0 102283, 2024

work page 2024

[3] [3]

D., Ho, J., Tarlow, D., and Van Den Berg, R

Austin, J., Johnson, D. D., Ho, J., Tarlow, D., and Van Den Berg, R. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34: 0 17981--17993, 2021

work page 2021

[4] [4]

and Albert, R

Barab \'a si, A.-L. and Albert, R. Emergence of scaling in random networks. science, 286 0 (5439): 0 509--512, 1999

work page 1999

[5] [5]

L., and Polozov, O

Brockschmidt, M., Allamanis, M., Gaunt, A. L., and Polozov, O. Generative code modeling with graphs. arXiv preprint arXiv:1805.08490, 2018

arXiv 2018

[6] [6]

Analog bits: Generating discrete data using diffusion models with self-conditioning

Chen, T., Zhang, R., and Hinton, G. Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202, 2022

arXiv 2022

[7] [7]

Efficient and degree-guided graph generation via discrete diffusion modeling

Chen, X., He, J., Han, X., and Liu, L.-P. Efficient and degree-guided graph generation via discrete diffusion modeling. arXiv preprint arXiv:2305.04111, 2023

arXiv 2023

[8] [8]

K., and Lu, X

Cherifi, H., Palla, G., Szymanski, B. K., and Lu, X. On community structure in complex networks: challenges and opportunities. Applied Network Science, 4 0 (1): 0 1--35, 2019

work page 2019