Pith. sign in

REVIEW 3 major objections 5 minor 65 references

FAROS: Fair Graph Generation via Attribute Switching Mechanisms

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read FAROS makes graph generators fairer by flipping node attributes mid-generation, without retraining.

desk verdict A promising retraining-free approach to fair graph generation, but the theoretical justification for its optimal node fraction rests on a strong assumption and the key hyperparameter selection is under-specified. read the letter →

arxiv 2507.03728 v1 pith:WACXLRH5 submitted 2025-07-04 cs.LG

Abdennacer Badaoui , Oussama Kharouiche , Hatim Mrabet , Daniele Malitesta , Fragkiskos D. Malliaros This is my paper
classification cs.LG
keywords fairgraphgenerationdiffusionmodelsattributeswitchinglinkpredictionstatisticalparityequalityofopportunityParetooptimality
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

FAROS is a method for making graph diffusion models generate fairer graphs: instead of retraining the generator with fairness constraints, it runs during generation and flips the sensitive attribute of a carefully chosen subset of nodes at a carefully chosen diffusion step. The paper claims this achieves a better accuracy-fairness trade-off on link prediction benchmarks, often beating retraining-based baselines and even reaching Pareto-optimal points in some settings. The authors care because diffusions are strong generators but amplify dataset bias, and prior fixes sacrifice accuracy or require retraining.

What carries the argument

The load-bearing machinery is the attribute switching mechanism: (1) a closed-form quadratic formula ρ* = arg min sign(R0)(R2ρ² + R1ρ + R0) that estimates the optimal fraction of nodes to switch by counteracting the imbalance between interior and exterior links, derived under an assumption that link existence depends only on sensitive attributes; (2) a multi-criteria selection of the switching time τ* that minimizes FGW(G, G̃) − γH(G̃), balancing the Fused Gromov-Wasserstein distance (node-topology distance, a proxy for accuracy) against an entropy of edge-attribute independence (a proxy for fairness).

What would settle it

On a citation or co-purchase network with rich non-sensitive features, compute the ρ* from the paper's quadratic formula and compare its accuracy-fairness frontier against a simple grid search over switching fractions: if the grid search finds a point that dominates ρ* in both accuracy and fairness, the closed-form optimality claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that fairness in graph generation can be improved by attribute switching alone: swapping the sensitive attribute of a computed optimal fraction ρ* of nodes at a computed optimal denoising step τ*, using a pre-trained graph diffusion model with no retraining. The tests on CORA, CITESEER, and AMAZON PHOTO show FAROS reduces statistical parity (ΔSP) and equality-of-opportunity (ΔEO) gaps while keeping AUC comparable or higher than several fairness-aware baselines, sometimes sitting on the Pareto frontier of accuracy vs. fairness.

Load-bearing premise

The closed-form optimal switching fraction is derived assuming that whether two nodes get linked depends only on their sensitive attributes, which is not true in real citation and co-purchase networks.

Editorial extensions

If this is right

  • A pre-trained graph generator can be made fairer without retraining, so the intervention applies when the generator is frozen or retraining is too costly.
  • The same attribute-switching approach can be mounted on other graph diffusion backbones; coupling FAROS with the asynchronous GraphMaker variant already gives strong trade-offs in the paper.
  • Because the method acts on sensitive attributes during generation, it could scale to large graphs where retraining with fairness constraints is infeasible, as long as the closed-form ρ* and FGW selection remain tractable.
  • The multi-criteria objective provides a principled way to tune the accuracy-fairness balance, instead of relying on ad-hoc early stopping or post-processing.
  • Fairer generated graphs can serve as training data for downstream tasks, so fairness improvements can propagate without modifying the downstream model.

Reading between the lines

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

  • Editorial inference: because the ρ* derivation assumes link existence depends only on sensitive attributes, the closed-form optimal fraction may fail on graphs where links are driven by other node features; the method might still work empirically, but the claimed optimality needs re-derivation or empirical search in those settings.
  • Editorial inference: the entropy-based fairness proxy encourages edge independence on sensitive attributes, which is closely related to statistical parity among generated edges; applying the same pipeline to other fairness notions would require modifying that objective.
  • Editorial inference: the approach could turn one biased generator into a family of generators indexed by ρ and τ, potentially enabling on-demand fairness control, though the paper does not explore this explicitly.
  • Editorial inference: since the method alters node attributes rather than topology, the fairness gain may come partly from relabeling rather than debiasing structural biases; a testable extension is to check whether switching preserves performance on tasks beyond link prediction, such as node classification.
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

3 major / 5 minor

Summary. FAROS proposes a training-free fairness intervention for pre-trained graph diffusion models (GDMs): at a chosen reverse-diffusion step τ*, a subset of nodes of size ρ* has its sensitive attribute resampled from a uniform or prior distribution, after which generation continues with the switched attributes. The fraction ρ* is derived from a quadratic optimization balancing expected interior/exterior edge counts, and τ* is selected by minimizing a multi-criteria objective combining a Fused Gromov-Wasserstein (FGW) distance as an accuracy proxy and an entropy term as a fairness proxy. On CORA, CITESEER, and AMAZON PHOTO, the authors report link-prediction results (AUC, ΔSP, ΔEO) showing that FAROS often improves fairness over GraphMaker while keeping accuracy close to, or above, several baselines, and that it is sometimes Pareto-optimal in the accuracy-fairness plane.

Significance. If the empirical claims hold, FAROS would be a valuable addition to the small literature on fairness in graph diffusion models: it requires no re-training of the GDM, is conceptually simple, and is evaluated against strong baselines on standard datasets with both uniform and prior-based switching variants. The paper also includes an analytic derivation of the switching fraction, an ablation of ρ* versus no-switching/all-switching, and an exploration of different diffusion backbones (GraphMaker-Sync, GraphMaker-Async, FairWire). These are useful contributions. However, both the derivation of ρ* and the selection of the balancing coefficient γ rest on assumptions or protocol choices that are not fully validated, and the Pareto-optimality evidence is therefore not yet conclusive.

major comments (3)
  1. [Supplementary Material, Section A.1 (Eq. (3))] The closed-form derivation of the optimal switching fraction ρ* relies on Assumption 2, that the existence of a link between two nodes depends only on their respective sensitive attributes. This assumption is explicitly stated in the supplement and is used to replace per-node-pair expectations with products of class counts. In citation and co-purchase networks, edge formation depends on node features and latent topics, not only on class labels, so the equality in Eq. (9) and the subsequent quadratic form in Eq. (16) are not justified. The authors should either (i) provide evidence that this approximation is accurate on the tested datasets, (ii) present ρ* as a heuristic rather than an optimal fraction, or (iii) derive the expression under a weaker assumption. As written, the optimality claim for ρ* is unsupported.
  2. [Supplementary Material, Section B.3 (Eq. (4))] The balancing coefficient γ is selected by grid search over {0.0, 0.1, ..., 1.0} and the paper reports that 'empirically finding a good balance' gives γ=0.5, but no validation split or selection protocol is described. Since γ directly controls the trade-off between the FGW accuracy proxy and the entropy fairness proxy in Eq. (4), and the same test metrics (AUC, ΔSP, ΔEO) are used both to choose γ and to rank methods, the Pareto-frontier comparisons in Figures 4-6 risk being circular. The authors should report how γ is chosen on a validation set (or show sensitivity of the reported results to γ), and they should state whether any test data were used during model selection.
  3. [Section 5, Conclusion] The conclusion states that FAROS 'achieves a better accuracy-fairness trade-off than other competitors' without the qualification 'in some tested settings' that appears in the abstract. On CITESEER, FairWire attains higher AUC (91.29±3.07) than both FAROS variants (84.45±7.71 and 87.76±6.21) while having worse fairness; whether this constitutes a 'better trade-off' depends on the decision maker's preference. The conclusion should be rephrased to match the more nuanced evidence actually presented.
minor comments (5)
  1. [Section 3.3] There is a typo in 'persfect edge independence' which should read 'perfect edge independence'.
  2. [Figure 2] The figure uses the symbol ω for what appears to be both ρ* and τ* in different subfigures, while the text consistently uses ρ* and τ*. Please harmonize the notation so that the pipeline diagram is readable.
  3. [Section 4.1 / Supplementary Material B.2] The paper states that results for FairAdj and Adversarial are taken directly from the FairWire paper rather than re-computed in the same pipeline. This should be stated more prominently in the main text, since a reader may otherwise assume all baselines were evaluated in a unified codebase.
  4. [Table 1 caption] The caption says 'Top-3 results are explicitly highlighted on each metric', but in the plain-text version there is no visible highlighting. Please use boldface, underlining, or superscript markers so the top-3 entries are identifiable.
  5. [Supplementary Material, Section B.3] The choice of the grid {1,2,3} for τ* in the synchronous GraphMaker setting follows GraphMaker's T=3, which is fine, but the statement that the grid-search 'steadily selected τ*=2' should be accompanied by the actual objective values for τ*=1 and τ*=3 (they are given in Figures 7-8, so a cross-reference would help).

Circularity Check

1 steps flagged · score 4.0 of 10

The headline accuracy-fairness trade-off is partly selected by fitting the balancing coefficient γ on the same datasets used to report the Pareto-frontier results.

  1. fitted input called prediction [Section 3.3, Eq. (4); Supplementary B.3]
    "τ ∗ = arg min τ ∈{T −1,...,1} FGW(G, G˜ρ∗,τ ) − γ H(G˜ρ∗,τ ). (4) ... Finally, within Equation (4), we also selected the coefficient γ (balancing the accuracy and fairness contributions) within {0.0, 0.1, . . . ,1.0}, empirically finding a good balance with γ = 0.5."

    The central empirical claim is that FAROS achieves a better accuracy-fairness trade-off than competitors (Abstract, Section 5, Table 1, Figures 4-6). In Eq. (4), γ directly controls the relative weight of the accuracy proxy (FGW) and the fairness proxy (H), and therefore determines which switching time τ ∗ is selected and thus the final reported (AUC, ΔSP, ΔEO) point. The paper states that γ was chosen by grid-search 'empirically finding a good balance', but it does not document a validation split or any independence between the γ selection and the test metrics used to construct the Pareto-frontier comparisons. With no such separation, the reported trade-off point is a selection outcome produced by the fitted γ rather than an independently predicted result.

full rationale

The ρ∗ computation in Eq. (3) and Supplementary A.1 is not itself circular: ρ∗ is computed from original-graph edge statistics and the switching distribution, not fitted to the reported ΔSP/ΔEO values; the downstream AUC/fairness numbers are obtained after training a GAE on generated graphs and testing on the original test set, so they are not equal by construction to the optimized FGW/entropy proxies. The self-citations in the paper are not load-bearing for the main derivation. The main circularity concern is the unvalidated selection of γ in Eq. (4): since γ balances the accuracy and fairness objectives before the reported trade-off is measured, and no validation split is described, the Pareto-frontier claim in Table 1 and Figures 4-6 is partly a selected outcome rather than an independent prediction. Supplementary A.1's Assumption 2 (link existence depends only on sensitive attributes) is a correctness risk for the closed-form ρ∗ in real networks, but it is an assumption, not a circular reduction of the claimed result.

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

The central mechanism rests on two explicit assumptions in Supplementary A.1 and two proxy assumptions in Section 3.3. The only fitted numeric parameter is gamma; alpha in the FGW distance is unreported. No invented entities are introduced.

free parameters (2)
  • gamma (γ) = 0.5
    Balances accuracy (FGW) and fairness (entropy) in Equation (4); selected empirically from {0.0,0.1,...,1.0} in B.3.
  • alpha (α) in FGW = not reported
    Balances Wasserstein and Gromov-Wasserstein terms in the FGW distance; no value or selection procedure is reported in the paper.
assumptions (4)
  • domain assumption Assumption 1: The pre-trained GDM perfectly mimics the original graph prior distribution of edges across sensitive attributes.
    Stated in Supplementary A.1; used to equate original and generated edge statistics in the derivation of rho*.
  • domain assumption Assumption 2: The existence of a link between two nodes depends only on their respective sensitive attributes.
    Stated in Supplementary A.1; this is load-bearing for the quadratic polynomial in rho. It is false for most real graphs.
  • ad hoc to paper Edge independence in the generated graph serves as a proxy for downstream link prediction fairness.
    Used in Section 3.3 to define the entropy objective; no evidence is provided that it correlates with dSP/dEO of a trained predictor.
  • ad hoc to paper Fused Gromov-Wasserstein distance serves as a proxy for downstream accuracy.
    Used in Equation (4) to preserve topology; the link to downstream AUC is assumed, not demonstrated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FAROS: Fair Graph Generation via Attribute Switching Mechanisms." pith.science (2026). https://pith.science/paper/WACXLRH5

@misc{pith2026250703728,
  author       = {Pith},
  title        = {Pith review of: FAROS: Fair Graph Generation via Attribute Switching Mechanisms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WACXLRH5}},
  note         = {Machine review of arXiv:2507.03728}
}
read the original abstract

Recent advancements in graph diffusion models (GDMs) have enabled the synthesis of realistic network structures, yet ensuring fairness in the generated data remains a critical challenge. Existing solutions attempt to mitigate bias by re-training the GDMs with ad-hoc fairness constraints. Conversely, with this work, we propose FAROS, a novel FAir graph geneRatiOn framework leveraging attribute Switching mechanisms and directly running in the generation process of the pre-trained GDM. Technically, our approach works by altering nodes' sensitive attributes during the generation. To this end, FAROS calculates the optimal fraction of switching nodes, and selects the diffusion step to perform the switch by setting tailored multi-criteria constraints to preserve the node-topology profile from the original distribution (a proxy for accuracy) while ensuring the edge independence on the sensitive attributes for the generated graph (a proxy for fairness). Our experiments on benchmark datasets for link prediction demonstrate that the proposed approach effectively reduces fairness discrepancies while maintaining comparable (or even higher) accuracy performance to other similar baselines. Noteworthy, FAROS is also able to strike a better accuracy-fairness trade-off than other competitors in some of the tested settings under the Pareto optimality concept, demonstrating the effectiveness of the imposed multi-criteria constraints.

Figures

Figures reproduced from arXiv: 2507.03728 by the authors.

Figure 1
Figure 1. Link prediction results in terms of accu [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall pipeline of FAROS. By starting from a pre-trained GDM [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Link prediction results for FAROS on CORA when switching no nodes (i.e., GraphMaker ), all nodes ( and ), and the optimal fraction of nodes ( and ). Accuracy is assessed using AUC, while fairness is measured in terms of ∆SP (a) and ∆EO (b). For better visualization, we computed 100 - value for each fairness metric to adhere to the principle “higher is better” [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Pareto visualization on CORA, with (a) AUC vs. ∆SP and (b) AUC vs. ∆EO. 85 90 Accuracy (%) 65 75 85 Fairness (%) (a) 85 90 Accuracy (%) 86 90 94 Fairness (%) (b) GraphMaker FairAdj Adversarial FairWire FAROS-Uniform FAROS-Prior Pareto-optimal Non Pareto-optimal [PITH_…
Figure 5
Figure 5. Figure 5: Pareto visualization on CITESEER, with (a) AUC vs. ∆SP and (b) AUC vs. ∆EO. 93 94 Accuracy (%) 70 72 74 Fairness (%) (a) 93 94 Accuracy (%) 88 92 96 Fairness (%) (b) GraphMaker Adversarial FairWire FAROS-Uniform FAROS-Prior Pareto-optimal Non Pareto-optimal [PITH_FULL…
Figure 6
Figure 6. Figure 6: Pareto visualization on AMAZON PHOTO, with (a) AUC vs. ∆SP and (b) AUC vs. ∆EO. Results for FairAdj was not computed on AMAZON PHOTO. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 7
Figure 7. Figure 7: Variation of the objective function (for [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]
Figure 8
Figure 8. Figure 8: Variation of the objective function (for [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

65 extracted references · 52 canonical work pages

  1. [1]

    Hamilton

    William L. Hamilton. Graph Representation Learning . Synthesis Lectures on Artificial Intelligence and Machine Learning. Morgan & Claypool Publishers, 2020

    work page 2020

  2. [2]

    Light- GCN: Simplifying and powering graph convolution network for recommendation

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yong-Dong Zhang, and Meng Wang. Light- GCN: Simplifying and powering graph convolution network for recommendation. In SIGIR, 2020

    work page 2020

  3. [3]

    LightGCL: Simple yet effective graph contrastive learning for recommendation

    Xuheng Cai, Chao Huang, Lianghao Xia, and Xubin Ren. LightGCL: Simple yet effective graph contrastive learning for recommendation. In ICLR, 2023

    work page 2023

  4. [4]

    Drug-target interaction pre- diction with graph regularized matrix factorization

    Ali Ezzat, Peilin Zhao, Min Wu, Xiaoli Li, and Chee Keong Kwoh. Drug-target interaction pre- diction with graph regularized matrix factorization. IEEE ACM Trans. Comput. Biol. Bioinform., 14(3):646–656, 2017

    work page 2017

  5. [5]

    DMHGNN: double multi-view heterogeneous graph neural network framework for drug-target interaction prediction

    Qiao Ning, Yue Wang, Yaomiao Zhao, Jiahao Sun, Lu Jiang, Kaidi Wang, and Minghao Yin. DMHGNN: double multi-view heterogeneous graph neural network framework for drug-target interaction prediction. Artif. Intell. Medicine, 159:103023, 2025

    work page 2025

  6. [6]

    Castro-Correa, Jhony H

    Jhon A. Castro-Correa, Jhony H. Giraldo, Mohsen Badiey, and Fragkiskos D. Malliaros. Gegen- bauer graph neural networks for time-varying signal reconstruction. IEEE Transactions on Neural Networks and Learning Systems , 35(9):11734–11745, 2024

    work page 2024

  7. [7]

    The graph neural network model

    Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Trans. Neural Networks, 20(1):61–80, 2009

    2009

  8. [8]

    Graph neural networks are inherently good generalizers: Insights by bridging GNNs and MLPs

    Chenxiao Yang, Qitian Wu, Jiahua Wang, and Junchi Yan. Graph neural networks are inherently good generalizers: Insights by bridging GNNs and MLPs. In ICLR, 2023

    work page 2023

Show all 65 references
  1. [9]

    Fairness in machine learning: A survey

    Simon Caton and Christian Haas. Fairness in machine learning: A survey. ACM Comput. Surv., 56(7):166:1–166:38, 2024

    2024

  2. [10]

    Say no to the discrimination: Learning fair graph neural networks with limited sensitive attribute information

    Enyan Dai and Suhang Wang. Say no to the discrimination: Learning fair graph neural networks with limited sensitive attribute information. In WSDM, 2021

    2021

  3. [11]

    On dyadic fairness: Exploring and mitigating bias in graph connections

    Peizhao Li, Yifei Wang, Han Zhao, Pengyu Hong, and Hongfu Liu. On dyadic fairness: Exploring and mitigating bias in graph connections. In ICLR, 2021

    2021

  4. [12]

    Fairdrop: Biased edge dropout for enhancing fairness in graph representation learning

    Indro Spinelli, Simone Scardapane, Amir Hussain, and Aurelio Uncini. Fairdrop: Biased edge dropout for enhancing fairness in graph representation learning. IEEE Trans. Artif. Intell., 3(3): 344–354, 2022

    2022

  5. [13]

    Fairgat: Fairness-aware graph attention networks

    Öykü Deniz Köse and Yanning Shen. Fairgat: Fairness-aware graph attention networks. ACM Trans. Knowl. Discov. Data, 18(7):164, 2024. 10

    2024

  6. [14]

    Fair graph U-Net: A fair graph learning framework integrating group and individual awareness

    Zichong Wang, Zhibo Chu, Thang Viet Doan, Shaowei Wang, Yongkai Wu, Vasile Palade, and Wenbin Zhang. Fair graph U-Net: A fair graph learning framework integrating group and individual awareness. In AAAI, 2025

    2025

  7. [15]

    FA*IR: A fair top-k ranking algorithm

    Meike Zehlike, Francesco Bonchi, Carlos Castillo, Sara Hajian, Mohamed Megahed, and Ricardo Baeza-Yates. FA*IR: A fair top-k ranking algorithm. In CIKM, 2017

    2017

  8. [16]

    Rossi, Namyong Park, Puja Trivedi, Yu Wang, Tong Yu, Sungchul Kim, Franck Dernoncourt, and Nesreen K

    April Chen, Ryan A. Rossi, Namyong Park, Puja Trivedi, Yu Wang, Tong Yu, Sungchul Kim, Franck Dernoncourt, and Nesreen K. Ahmed. Fairness-aware graph neural networks: A survey. ACM Trans. Knowl. Discov. Data, 18(6):138:1–138:23, 2024

    2024

  9. [17]

    A survey on fairness for machine learning on graphs

    Manvi Choudhary, Charlotte Laclau, and Christine Largeron. A survey on fairness for machine learning on graphs. CoRR, abs/2205.05396, 2022

    2022 arXiv

  10. [18]

    Fairness in graph mining: A survey

    Yushun Dong, Jing Ma, Song Wang, Chen Chen, and Jundong Li. Fairness in graph mining: A survey. IEEE Trans. Knowl. Data Eng., 35(10):10583–10602, 2023

    2023

  11. [19]

    Hanqun Cao, Cheng Tan, Zhangyang Gao, Yilun Xu, Guangyong Chen, Pheng-Ann Heng, and Stan Z. Li. A survey on generative diffusion models. IEEE Trans. Knowl. Data Eng. , 36(7): 2814–2830, 2024

    2024

  12. [20]

    A survey on video diffusion models

    Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. ACM Comput. Surv., 57(2):41:1–41:42, 2025

    2025

  13. [21]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014

    2014

  14. [22]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. CoRR, abs/1406.2661, 2014

    2014 arXiv

  15. [23]

    Compre- hensive exploration of diffusion models in image generation: a survey

    Hang Chen, Qian Xiang, Jiaxin Hu, Meilin Ye, Chao Yu, Hao Cheng, and Lei Zhang. Compre- hensive exploration of diffusion models in image generation: a survey. Artif. Intell. Rev., 58(4): 99, 2025

    2025

  16. [24]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021

    2021

  17. [25]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022

    2022

  18. [26]

    Very deep V AEs generalize autoregressive models and can outperform them on images

    Rewon Child. Very deep V AEs generalize autoregressive models and can outperform them on images. In ICLR, 2021

    2021

  19. [27]

    Fleet, and Mohammad Norouzi

    Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Trans. Pattern Anal. Mach. Intell., 45(4):4713–4726, 2023

    2023

  20. [28]

    SSD-LM: semi-autoregressive simplex- based diffusion language model for text generation and modular control

    Xiaochuang Han, Sachin Kumar, and Yulia Tsvetkov. SSD-LM: semi-autoregressive simplex- based diffusion language model for text generation and modular control. In ACL, 2023

    2023

  21. [29]

    Hashimoto

    Xiang Lisa Li, John Thickstun, Ishaan Gulrajani, Percy Liang, and Tatsunori B. Hashimoto. Diffusion-lm improves controllable text generation. In NeurIPS, 2022

    2022

  22. [30]

    Hierarchical text-conditional image generation with CLIP latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022

    2022 arXiv

  23. [31]

    Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal LLMs

    Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and Bin Cui. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal LLMs. In ICML, 2024

    2024

  24. [32]

    Generative diffusion models on graphs: Methods and applications

    Chengyi Liu, Wenqi Fan, Yunqing Liu, Jiatong Li, Hang Li, Hui Liu, Jiliang Tang, and Qing Li. Generative diffusion models on graphs: Methods and applications. In IJCAI, 2023

    2023

  25. [33]

    A survey on deep graph generation: Methods and applications

    Yanqiao Zhu, Yuanqi Du, Yinkai Wang, Yichen Xu, Jieyu Zhang, Qiang Liu, and Shu Wu. A survey on deep graph generation: Methods and applications. In LoG, 2022. 11

    2022

  26. [34]

    A systematic survey on deep generative models for graph generation

    Xiaojie Guo and Liang Zhao. A systematic survey on deep generative models for graph generation. IEEE Trans. Pattern Anal. Mach. Intell., 45(5):5370–5390, 2023

    2023

  27. [35]

    Equivariant diffusion for molecule generation in 3d

    Emiel Hoogeboom, Victor Garcia Satorras, Clément Vignac, and Max Welling. Equivariant diffusion for molecule generation in 3d. In ICML, 2022

    2022

  28. [36]

    GeoDiff: A geometric diffusion model for molecular conformation generation

    Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. GeoDiff: A geometric diffusion model for molecular conformation generation. In ICLR, 2022

    2022

  29. [37]

    A survey on graph diffusion models: Generative AI in science for molecule, protein and material

    Mengchun Zhang, Maryam Qamar, Taegoo Kang, Yuna Jung, Chenshuang Zhang, Sung-Ho Bae, and Chaoning Zhang. A survey on graph diffusion models: Generative AI in science for molecule, protein and material. CoRR, abs/2304.01565, 2023

    2023 arXiv

  30. [38]

    Directional diffusion models for graph representation learning

    Run Yang, Yuling Yang, Fan Zhou, and Qiang Sun. Directional diffusion models for graph representation learning. In NeurIPS, 2023

    2023

  31. [39]

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

    Jaehyeong Jo, Seul Lee, and Sung Ju Hwang. Score-based generative modeling of graphs via the system of stochastic differential equations. In ICML, 2022

    2022

  32. [40]

    Permutation invariant graph generation via score-based generative modeling

    Chenhao Niu, Yang Song, Jiaming Song, Shengjia Zhao, Aditya Grover, and Stefano Ermon. Permutation invariant graph generation via score-based generative modeling. In AISTATS, 2020

    2020

  33. [41]

    GraphGDP: Generative diffusion processes for permutation invariant graph generation

    Han Huang, Leilei Sun, Bowen Du, Yanjie Fu, and Weifeng Lv. GraphGDP: Generative diffusion processes for permutation invariant graph generation. In ICDM, pages 201–210. IEEE, 2022

    2022

  34. [42]

    Digress: Discrete denoising diffusion for graph generation

    Clément Vignac, Igor Krawczuk, Antoine Siraudin, Bohan Wang, V olkan Cevher, and Pascal Frossard. Digress: Discrete denoising diffusion for graph generation. In ICLR, 2023

    2023

  35. [43]

    Discrete-state continuous-time diffusion for graph generation

    Zhe Xu, Ruizhong Qiu, Yuzhong Chen, Huiyuan Chen, Xiran Fan, Menghai Pan, Zhichen Zeng, Mahashweta Das, and Hanghang Tong. Discrete-state continuous-time diffusion for graph generation. In NeurIPS, 2024

    2024

  36. [44]

    Inverse molecular design with multi- conditional diffusion guidance

    Gang Liu, Jiaxin Xu, Tengfei Luo, and Meng Jiang. Inverse molecular design with multi- conditional diffusion guidance. CoRR, abs/2401.13858, 2024

    2024 arXiv

  37. [45]

    Malliaros, and Christopher Morris

    Antoine Siraudin, Fragkiskos D. Malliaros, and Christopher Morris. Cometh: A continuous-time discrete-state graph diffusion model. Transactions on Machine Learning Research, 2025

    2025

  38. [46]

    Potluru, and Pan Li

    Mufei Li, Eleonora Kreacic, Vamsi K. Potluru, and Pan Li. GraphMaker: Can diffusion models generate large attributed graphs? Transactions on Machine Learning Research, 2024

    2024

  39. [47]

    Rossi, David Arbour, Tong Yu, Franck Dernoncourt, Sungchul Kim, Nedim Lipka, Namyong Park, Nesreen K

    Puja Trivedi, Ryan A. Rossi, David Arbour, Tong Yu, Franck Dernoncourt, Sungchul Kim, Nedim Lipka, Namyong Park, Nesreen K. Ahmed, and Danai Koutra. Editing partially observ- able networks via graph diffusion models. In ICML, 2024

    2024

  40. [48]

    Fair diffusion: Instructing text-to-image generation models on fairness

    Felix Friedrich, Patrick Schramowski, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Sasha Luccioni, and Kristian Kersting. Fair diffusion: Instructing text-to-image generation models on fairness. CoRR, abs/2302.10893, 2023

    2023 arXiv

  41. [49]

    Easily accessible text-to- image generation amplifies demographic stereotypes at large scale

    Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to- image generation amplifies demographic stereotypes at large scale. In F AccT, 2023

    2023

  42. [50]

    DALL-EV AL: probing the reasoning skills and social biases of text-to-image generation models

    Jaemin Cho, Abhay Zala, and Mohit Bansal. DALL-EV AL: probing the reasoning skills and social biases of text-to-image generation models. In ICCV, 2023

    2023

  43. [51]

    Malliaros, Mirko Marras, and Ernesto William De Luca

    Daniele Malitesta, Giacomo Medda, Erasmo Purificato, Ludovico Boratto, Fragkiskos D. Malliaros, Mirko Marras, and Ernesto William De Luca. How fair is your diffusion recommender model? CoRR, abs/2409.04339, 2024

    2024 arXiv

  44. [52]

    Fair sampling in diffusion models through switching mechanism

    Yujin Choi, Jinseong Park, Hoki Kim, Jaewook Lee, and Saerom Park. Fair sampling in diffusion models through switching mechanism. In AAAI, 2024. 12

    2024

  45. [53]

    Kankanhalli

    Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan S. Kankanhalli. Finetuning text-to-image diffusion models for fairness. CoRR, abs/2311.07604, 2023

    2023 arXiv

  46. [54]

    FairGen: Controlling sensi- tive attributes for fair generations in diffusion models via adaptive latent guidance

    Mintong Kang, Vinayshekhar Bannihatti Kumar, Shamik Roy, Abhishek Kumar, Sopan Khosla, Balakrishnan Narayanaswamy, and Rashmi Gangadharaiah. FairGen: Controlling sensi- tive attributes for fair generations in diffusion models via adaptive latent guidance. CoRR, abs/2503.01872, 2025

    2025

  47. [55]

    Marques, and Santiago Segarra

    Madeline Navarro, Samuel Rey, Andrei Buciulea, Antonio G. Marques, and Santiago Segarra. Mitigating subpopulation bias for fair network topology inference. In EUSIPCO, 2024

    2024

  48. [56]

    Marques, and Santiago Segarra

    Madeline Navarro, Samuel Rey, Andrei Buciulea, Antonio G. Marques, and Santiago Segarra. Fair GLASSO: estimating fair graphical models with unbiased statistical behavior. In NeurIPS, 2024

    2024

  49. [57]

    FairGen: Towards fair graph generation

    Lecheng Zheng, Dawei Zhou, Hanghang Tong, Jiejun Xu, Yada Zhu, and Jingrui He. FairGen: Towards fair graph generation. In ICDE, 2024

    2024

  50. [58]

    FairWire: Fair graph generation

    Oyku Deniz Kose and Yanning Shen. FairWire: Fair graph generation. In NeurIPS, 2024

    2024

  51. [59]

    Optimal transport for structured data with application on graphs

    Titouan Vayer, Nicolas Courty, Romain Tavenard, Laetitia Chapel, and Rémi Flamary. Optimal transport for structured data with application on graphs. In ICML, 2019

    2019

  52. [60]

    Collective classification in network data

    Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Gallagher, and Tina Eliassi- Rad. Collective classification in network data. AI Mag., 29(3):93–106, 2008

    2008

  53. [61]

    Pitfalls of graph neural network evaluation

    Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. Pitfalls of graph neural network evaluation. CoRR, abs/1811.05868, 2018

    2018 arXiv

  54. [62]

    Sagess: A sampling graph denoising diffusion model for scalable graph generation

    Stratis Limnios, Praveen Selvaraj, Mihai Cucuringu, Carsten Maple, Gesine Reinert, and Andrew Elliott. Sagess: A sampling graph denoising diffusion model for scalable graph generation. In ECAI, 2024

    2024

  55. [63]

    Optimal Transport: Old and New

    Cédric Villani. Optimal Transport: Old and New . Grundlehren der mathematischen Wis- senschaften. Springer Berlin, Heidelberg, 2008

    2008

  56. [64]

    Gromov-wasserstein distances and the metric approach to object matching

    Facundo Mémoli. Gromov-wasserstein distances and the metric approach to object matching. F ound. Comput. Math., 11(4):417–487, 2011

    2011

  57. [65]

    original graph

    Justin Solomon, Gabriel Peyré, Vladimir G. Kim, and Suvrit Sra. Entropic metric alignment for correspondence problems. ACM Trans. Graph., 35(4):72:1–72:13, 2016. 13 Supplementary Material The Supplementary Material is organized as follows. First, in Appendix A, we provide the ...

    2016

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.