Pith. sign in

REVIEW 5 minor 46 references

A U-shaped GNN reverse diffusion process samples stochastic graph signals from unknown conditional laws given topology and node features.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-10 20:21 UTC pith:QAH5A7JE

load-bearing objection Solid engineering paper: a clean U-GNN that keeps graph filtering on the original topology, with thorough generative results on two real tasks and released code.

arxiv 2607.06833 v1 pith:QAH5A7JE submitted 2026-07-07 cs.LG eess.SP

Generative Diffusion Models of Stochastic Graph Signals

Yi\u{g}it Berkay Uslu , Samar Hadou , Sergio Rozada , Shirin Saeedi Bidokhti , Alejandro Ribeiro This is my paper
classification cs.LG eess.SP
keywords diffusion modelsgraph signalsgraph neural networksU-GNNwireless resource allocationfinancial forecastingconditional generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Many graph machine-learning tasks—stock forecasting, wireless resource allocation, recommender systems—need samples from unknown conditional distributions of signals living on graphs, not just a single predicted average. Prevailing designs are application-specific and often collapse to conditional means. This paper unifies those problems as conditional graph-signal generative modeling and solves them with one denoising diffusion framework whose reverse process is a U-Graph Neural Network. The U-GNN adapts the image U-Net by treating pooling as learned nested node selection with zero-padded lifting and by running strided graph convolutions on the original graph, so multi-resolution filtering stays convolutional without explicit coarsening. On S&P 500 return trajectories and multi-user wireless power control, the sampler matches or approaches reference and expert distributions while capturing uncertainty and rare events that mean-based methods miss.

Core claim

A reverse diffusion process parametrized by a U-GNN draws samples from the conditional law of graph signals given the graph shift operator and node-feature side information, realizing a single generative model that works for both stock-price forecasting and optimal wireless resource allocation without task-tailored architectures that regress to a conditional mean.

What carries the argument

The U-Graph Neural Network (U-GNN): multi-resolution encoder–decoder GNN blocks whose pooling/unpooling are learned nested selection matrices plus zero-padded lifting, while strided convolutions on the original graph set hop reach and keep every resolution convolutional without building coarsened graphs.

Load-bearing premise

Learning which nodes to keep at each resolution, then filtering with strides on the original graph, is assumed to preserve enough multi-scale structure for the reverse process to recover the full conditional distribution rather than collapsing to a mean.

What would settle it

On held-out S&P 500 windows or wireless networks, U-GNN samples fail to match reference distributions on both accuracy metrics (CRPS, rate percentiles) and distributional diagnostics (return autocorrelation, covariance eigenvalues, time-sharing feasibility) while a mean-regression baseline succeeds.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Stock-return ensembles improve CRPS, RMSE and stylized-fact fidelity over a geometric random-walk baseline.
  • Near-optimal wireless power allocations can be drawn in one accelerated DDIM pass that approaches primal–dual expert cell-edge rates and feasibility without online optimization.
  • The same backbone transfers across network sizes without retraining, consistent with GNN equivariance.
  • Generative time-sharing sustains QoS feasibility that deterministic mean allocations lose.
  • Graph-aware or latent diffusion and graph-transformer hybrids become direct extensions of the same reverse-process backbone.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same reverse process could serve as a drop-in generative policy for other offline-optimized graph problems (routing, ranking) where an expert supplies training samples.
  • Because filtering never leaves the original graph, transfer across topologies may be cleaner than methods that rebuild coarsened graphs at every resolution.
  • The observed overshoot of excess kurtosis in finance points to a tunable fidelity–sharpness trade-off controllable by checkpoint selection or sampling temperature.
  • Adding long-range temporal attention, as the authors flag, would test whether multi-resolution node selection alone is sufficient for strongly sequential graph signals.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

0 major / 5 minor

Summary. The paper unifies conditional generation of stochastic graph signals under a denoising diffusion framework. A reverse process is parametrized by a U-Graph Neural Network (U-GNN) that generalizes the U-Net via nested selection matrices for learned node (un)pooling, zero-padded lifting, and strided graph convolutions that operate on the original GSO without explicit coarsening (Secs. II–IV, Eqs. (14)–(25)). Conditioning is on topology S and node features u. The method is demonstrated on S&P 500 multi-step return forecasting and wireless power allocation under ergodic QoS constraints, with extensive held-out metrics (Tables II, IV; Figs. 3–6) showing improved distributional fidelity over a geometric random walk and near-expert cell-edge rates/feasibility via multi-modal time-sharing.

Significance. If the empirical claims hold, the work supplies a reusable, architecture-level solution for sampling from unknown conditional laws of graph signals rather than regressing to conditional means. The U-GNN construction (nested D_b, lift–filter–reduce, stride) is fully specified, code is released, and the two applications are non-trivial and complementary: one recovers stylized facts and score-based NLL distributions on financial trajectories, the other realizes expert-level feasibility through stochastic time-sharing that a mean allocation cannot achieve and transfers across network sizes. These are concrete engineering contributions for graph ML, wireless optimization, and spatio-temporal forecasting.

minor comments (5)
  1. The GRW baseline in Sec. V-A is deliberately weak (i.i.d. per-stock Gaussians). A short comparison to a stronger spatio-temporal baseline (e.g., a graph-aware AR or simple GNN forecaster) would better situate the gains in Table II and Fig. 4, even if only as an ablation.
  2. Fig. 2 caption and the surrounding text refer to both “Fig. 1b” and “Fig. 2”; the numbering of the architecture schematic should be made consistent with the main-text figure list.
  3. Appendix D discusses classical sum/max-pooling but reports that it added overhead without gain. A one-sentence statement of this negative result in the main text (near Remark 1 or Sec. IV-B) would help readers who expect neighborhood pooling.
  4. Notation for the composite selection matrices D_b (Eq. (18)) and the per-level C_b is clear, but a brief reminder that D_1 = I_N appears only once; repeating it when the encoder equations (19)–(21) are introduced would reduce lookup.
  5. In the WRA setup, the expert is trained per density group and supplies the training targets. Clarifying that the same expert is run on held-out networks solely for evaluation (already implied) would forestall any residual concern about train–test leakage.

Circularity Check

0 steps flagged

No significant circularity: standard conditional diffusion trained on external data/expert samples, evaluated on held-out splits against independent baselines.

full rationale

The paper's derivation chain is ordinary supervised denoising diffusion (noise-prediction ELBO on reparameterized forward process, reverse kernels via U-GNN) conditioned on given (S, u). Training targets are external: historical S&P 500 returns (sliding windows, interleaved chronological splits) and allocations collected from a separately trained primal-dual expert GNN (not derived inside the diffusion). Evaluation uses held-out windows/networks, CRPS/MIS/RMSE, stylized-fact gaps, Wasserstein on NLL/ELBO scores, ergodic-rate percentiles/feasibility, and size-transfer, versus GRW, full-power, average-power, and the expert itself. Self-citations [1],[2] are explicitly labeled preliminary conference/preprint versions of the same line of work and are not invoked as uniqueness theorems or load-bearing external facts that force the present claims. Nested selection matrices, lift-filter-reduce layers, and stride are constructive architectural definitions (Secs. III-IV, Apps. B-C), not tautological rewritings of the generative fidelity claim. No equation reduces a reported prediction or distributional match to a fitted constant by construction. The work is self-contained engineering of a GNN U-Net backbone for graph-signal diffusion.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The paper is a methodological contribution resting on standard diffusion and GNN theory plus a small set of architectural design choices and hyperparameters. No new physical entities are postulated; free parameters are ordinary neural-network and diffusion knobs fitted or chosen for the two tasks.

free parameters (5)
  • U-GNN depth B and pooling factors ρ = B=4, ρ=2
    Fixed to B=4, ρ=2 (overall 8× reduction); chosen by design, not derived.
  • Channel widths F_b, GNN layers L, hops K, stride γ_max = 64, 2, 2, 2
    Set to 64 / 2 / 2 / 2; standard capacity knobs.
  • Diffusion steps K, β schedule, DDIM steps and η = 500 / linear / 100 / 0.2
    K=500 linear β, 100 DDIM steps η=0.2; conventional choices.
  • Selector temperature τ and Gumbel exploration ε schedules = linear anneal to floors
    Annealed from (1,1) to (0.5,0) over training; STE hyperparameters.
  • RevIN blend weight and scale (S&P only) = 0.7 / 1.107
    0.7 and 1.107; task-specific normalization constants.
axioms (4)
  • standard math Standard DDPM/DDIM forward process and noise-prediction ELBO equivalence (Ho et al., Song et al.)
    Used throughout Sec. II; no re-derivation required.
  • domain assumption Graph-shift equivariance and stability properties of polynomial GNN filters
    Invoked for transferability claims (Sec. I, V-B, Fig. 6); standard GNN theory.
  • ad hoc to paper Learned Top-K node selection with straight-through estimator yields a usable multi-resolution hierarchy
    Core design choice of U-GNN (Sec. IV-B, App. C); not proved to be optimal, only shown empirically.
  • domain assumption Expert primal–dual iterates supply near-optimal samples from the true optimal allocation distribution
    WRA supervision (Sec. V-B); quality of the generative target depends on this.
invented entities (2)
  • U-Graph Neural Network (U-GNN) no independent evidence
    purpose: Multi-resolution encoder–decoder denoiser that generalizes U-Net to graph signals via nested selection matrices and strided lift–filter–reduce layers.
    Central architectural invention; independent evidence is the empirical performance on two tasks and the public code, not an external physical prediction.
  • Nested selection matrices D_b = product of C_j with zero-padded lifting no independent evidence
    purpose: Realize pooling/unpooling while keeping all convolutions on the original graph.
    Technical device introduced in Sec. III–IV; no external falsifiable handle beyond the paper’s experiments.

pith-pipeline@v1.1.0-grok45 · 33509 in / 3158 out tokens · 39044 ms · 2026-07-10T20:21:33.228364+00:00 · methodology

0 comments
read the original abstract

Sampling stochastic signals supported on a graph underlies many graph machine learning tasks, including recommender systems, forecasting in financial markets, and wireless network optimization. In these settings, the target signals are realizations of unknown conditional distributions. However, prevailing approaches rely mostly on intricate, application-tailored designs that often regress to a conditional mean instead of sampling from the conditional law. This paper unifies such problems as conditional graph signal generative modeling and tackles them with a single denoising diffusion framework. We learn a reverse diffusion process, parametrized by graph neural networks (GNNs), that draws graph signals conditioned directly on the graph topology and on node-feature side information. The reverse process is realized by a novel architecture, the U-Graph Neural Network (U-GNN), which generalizes the image-convolutional U-Net to graph-structured signals. The U-GNN performs multi-resolution encoder--decoder processing in which pooling and unpooling reduce to a learned node selection, expressed by nested selection matrices, and a zero-padded lifting of coarse signals back to the full node set. The graph convolutions are carried out on the original graph, with a stride that sets their hop reach, so the U-GNN bypasses explicit graph coarsening at every resolution. We demonstrate our method on two generative tasks: stock price forecasting and optimal wireless resource allocation, with extensive numerical results in both domains.

Figures

Figures reproduced from arXiv: 2607.06833 by Alejandro Ribeiro, Samar Hadou, Sergio Rozada, Shirin Saeedi Bidokhti, Yi\u{g}it Berkay Uslu.

Figure 1
Figure 1. Figure 1: A denoising diffusion model of graph signals. A forward noising process q(xk | xk−1) gradually removes structure from the original graph signals by adding white noise, with xK ≈ N (0, I). A denoising diffusion process pθ(xk−1|xk; S, u) is trained to reverse the forward process and generate novel graph signal samples x0 distributed approximately by the conditional distributions qdata(x0 | S, u) for a given … view at source ↗
Figure 2
Figure 2. Figure 2: U-GNN denoiser architecture. A U-GNN of B=4 depths maps the noisy graph signal xk to a noise estimate ϵθ. At each depth, a projection Πb (purple) fuses the global embeddings [U0; K0] of node states u and diffusion step k into the signal path before a GNN module Φ E/D b (green), and decoder blocks additionally merge the encoder skips. Learned selectors Ψb (orange) produce selection matrices Cb+1 that down-s… view at source ↗
Figure 3
Figure 3. Figure 3: Example S&P 500 forecasting trajectories from U-GNN. Each column is a different test window with three panels for arbitrarily chosen stocks. In each panel, the solid black line is the observed history whereas the solid blue line is the ground-truth future, the thin orange lines are individual trajectories sampled from the diffusion ensemble, and the solid orange line is their mean, over the forecast horizo… view at source ↗
Figure 4
Figure 4. Figure 4: Distributional fidelity diagnostics on the S&P 500 test split. (a) Temporal structure: autocorrelation of returns r and squared returns r 2 for real and generated series, probing linear predictability and volatility clustering. (b) Spatial (cross￾sectional) structure: the top-5 eigenvalues of the return covariance matrix Cˆ, capturing how well each model reproduces inter￾stock correlation. (c)–(d) Histogra… view at source ↗
Figure 3
Figure 3. Figure 3: The sampled trajectories spread out from the last [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: U-GNN attains ergodic feasibility through time-sharing. We zoom in on an example test network, where each node is a tx–rx pair and edges indicate interference strength. The top row shows the per-node normalized transmit power p/Pmax (logarithmic color scale) for three consecutive slots of the sequential allocation sampled by the U-GNN policy, alongside the mean allocation E[p] (rightmost). The bottom row s… view at source ↗
Figure 6
Figure 6. Figure 6: Size-transferability of the U-GNN policy. Ergodic 5th-percentile (p5) rate vs. network size for four interference densities (tx–rx pairs/km2 ) at fmin = 0.6 bits/s/Hz (red dashed line). Diamonds mark the native training size (400 pairs), and all other sizes are evaluated without retraining. receiver. Crucially, this rate averages the per-slot rates and must not be conflated with the rate of the single mean… view at source ↗
Figure 2
Figure 2. Figure 2: Block interface (wide), oriented as Fig. 1 leftthe chain Π→ Φ E → split runs along the Block interface of a U-GNN encoder–decoder pair. [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 46 canonical work pages · 2 internal anchors

  1. [1]

    Graph signal generative diffusion 12 models,

    Yi ˘git Berkay Uslu, Samar Hadou, Sergio Rozada, Shirin Saeedi Bidokhti, and Alejandro Ribeiro, “Graph signal generative diffusion 12 models,” inIEEE Intl. Conf. Acoust., Speech and Signal Process. (ICASSP). IEEE, 2026, pp. 626–630

    work page 2026

  2. [2]

    Graph Signal Diffusion Models for Wireless Resource Allocation

    Yigit Berkay Uslu, Samar Hadou, Shirin Saeedi Bidokhti, and Alejandro Ribeiro, “Graph signal diffusion models for wireless resource alloca- tion,”arXiv preprint arXiv:2604.05175, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  3. [3]

    Neural graph collaborative filtering,

    Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua, “Neural graph collaborative filtering,” inIntl. ACM SIGIR Conf. Research and Develop. Info. Retrieval, 2019, pp. 165–174

    work page 2019

  4. [4]

    Ultragcn: Ultra simplification of graph convolutional networks for recommendation,

    Kelong Mao, Jieming Zhu, Xi Xiao, Biao Lu, Zhaowei Wang, and Xiuqiang He, “Ultragcn: Ultra simplification of graph convolutional networks for recommendation,” inACM Intl. Conf. Info. and Knowledge Management (CIKM), 2021, pp. 1253–1262

    work page 2021

  5. [5]

    Personalized graph signal processing for collaborative filtering,

    Jiahao Liu, Dongsheng Li, Hansu Gu, Tun Lu, Peng Zhang, Li Shang, and Ning Gu, “Personalized graph signal processing for collaborative filtering,” inACM Web Conf. (WWW), 2023, pp. 1264–1272

    work page 2023

  6. [6]

    Optimal wireless resource allo- cation with random edge graph neural networks,

    Mark Eisen and Alejandro Ribeiro, “Optimal wireless resource allo- cation with random edge graph neural networks,”IEEE Trans. Signal Process., vol. 68, pp. 2977–2991, 2020

    work page 2020

  7. [7]

    Graph neural networks for wireless communications: From theory to practice,

    Yifei Shen, Jun Zhang, S. H. Song, and Khaled B. Letaief, “Graph neural networks for wireless communications: From theory to practice,” IEEE Trans. Wireless Commun., vol. 22, no. 5, pp. 3554–3569, 2023

    work page 2023

  8. [8]

    Link scheduling using graph neural networks,

    Zhongyuan Zhao, Gunjan Verma, Chirag Rao, Ananthram Swami, and Santiago Segarra, “Link scheduling using graph neural networks,”IEEE Trans. Wireless Commun., vol. 22, no. 6, pp. 3997–4012, 2023

    work page 2023

  9. [9]

    ENGNN: A general edge-update empowered GNN architecture for radio resource management in wireless networks,

    Yunqi Wang, Yang Li, Qingjiang Shi, and Yik-Chung Wu, “ENGNN: A general edge-update empowered GNN architecture for radio resource management in wireless networks,”IEEE Trans. Wireless Commun., vol. 23, no. 6, pp. 5330–5344, 2024

    work page 2024

  10. [10]

    Deep graph unfolding for beamforming in mu-mimo inter- ference networks,

    Arindam Chowdhury, Gunjan Verma, Ananthram Swami, and Santiago Segarra, “Deep graph unfolding for beamforming in mu-mimo inter- ference networks,”IEEE Trans. Wireless Commun., vol. 23, no. 5, pp. 4889–4903, 2024

    work page 2024

  11. [11]

    Fast state-augmented learning for wireless resource allocation with dual variable regression,

    Yi ˘git Berkay Uslu, Navid NaderiAlizadeh, Mark Eisen, and Alejandro Ribeiro, “Fast state-augmented learning for wireless resource allocation with dual variable regression,” 2025

    work page 2025

  12. [12]

    Incorporating corporation relationship via graph convolutional neural networks for stock price prediction,

    Yingmei Chen, Zhongyu Wei, and Xuanjing Huang, “Incorporating corporation relationship via graph convolutional neural networks for stock price prediction,” inACM Intl. Conf. Info. and Knowledge Management (CIKM), 2018, pp. 1655–1658

    work page 2018

  13. [13]

    Spatiotemporal hypergraph convolution network for stock movement forecasting,

    Ramit Sawhney, Shivam Agarwal, Arnav Wadhwa, and Rajiv Ratn Shah, “Spatiotemporal hypergraph convolution network for stock movement forecasting,” inIEEE Intl. Conf. Data Mining, 2020, pp. 482–491

    work page 2020

  14. [14]

    Attention based dynamic graph neural network for asset pricing,

    Ajim Uddin, Xinyuan Tao, and Dantong Yu, “Attention based dynamic graph neural network for asset pricing,”Global Finance Journal, vol. 58, 2023

    work page 2023

  15. [15]

    Stationary signal processing on graphs,

    Nathanael Perraudin and Pierre Vandergheynst, “Stationary signal processing on graphs,”IEEE Trans. Signal Process., vol. 65, no. 13, pp. 3462–3477, 2017

    work page 2017

  16. [16]

    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,” inIntl. Conf. on Machine Learning (ICML), 2022, vol. 162 ofPMLR, pp. 10362–10383

    work page 2022

  17. [17]

    DiGress: Discrete denoising diffusion for graph generation,

    Cl ´ement Vignac, Igor Krawczuk, Antoine Siraudin, Bohan Wang, V olkan Cevher, and Pascal Frossard, “DiGress: Discrete denoising diffusion for graph generation,” inIntl. Conf. on Learning Repr. (ICLR), 2023

    work page 2023

  18. [18]

    Equivariant diffusion for molecule generation in 3d,

    Emiel Hoogeboom, V ´ıctor Garcia Satorras, Cl ´ement Vignac, and Max Welling, “Equivariant diffusion for molecule generation in 3d,” inIntl. Conf. on Machine Learning (ICML), 2022, vol. 162 ofPMLR, pp. 8867– 8887

    work page 2022

  19. [19]

    Auto-encoding variational bayes,

    Diederik P. Kingma and Max Welling, “Auto-encoding variational bayes,” inIntl. Conf. on Learning Repr. (ICLR), 2014

    work page 2014

  20. [20]

    Generative adversarial nets,

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio, “Generative adversarial nets,” inAdvances in Neural Info. Process. Syst., 2014, vol. 27, pp. 2672–2680

    work page 2014

  21. [21]

    Variational inference with normalizing flows,

    Danilo Jimenez Rezende and Shakir Mohamed, “Variational inference with normalizing flows,” inIntl. Conf. on Machine Learning (ICML), 2015, vol. 37 ofPMLR, pp. 1530–1538

    work page 2015

  22. [22]

    Denoising diffusion probabilistic models,

    Jonathan Ho, Ajay Jain, and Pieter Abbeel, “Denoising diffusion probabilistic models,” inAdvances in Neural Info. Process. Syst., 2020, vol. 33, pp. 6840–6851

    work page 2020

  23. [23]

    Denoising diffusion implicit models,

    Jiaming Song, Chenlin Meng, and Stefano Ermon, “Denoising diffusion implicit models,” inIntl. Conf. on Learning Repr. (ICLR), 2021

    work page 2021

  24. [24]

    Score-based generative modeling through stochastic differential equations,

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Ku- mar, Stefano Ermon, and Ben Poole, “Score-based generative modeling through stochastic differential equations,” inIntl. Conf. on Learning Repr. (ICLR), 2021

    work page 2021

  25. [25]

    Flow matching for generative modeling,

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le, “Flow matching for generative modeling,” inIntl. Conf. on Learning Repr. (ICLR), 2023

    work page 2023

  26. [26]

    DiffSTG: Probabilistic spatio- temporal graph forecasting with denoising diffusion models,

    Haomin Wen, Youfang Lin, Yutong Xia, Huaiyu Wan, Qingsong Wen, Roger Zimmermann, and Yuxuan Liang, “DiffSTG: Probabilistic spatio- temporal graph forecasting with denoising diffusion models,” inACM Intl. Conf. Adv. Geo. Info. Syst. (SIGSPATIAL). 2023, pp. 1–12, ACM

    work page 2023

  27. [27]

    DiffSTOCK: Probabilistic relational stock market predictions using diffusion models,

    Divyanshu Daiya, Monika Yadav, and Harshit Singh Rao, “DiffSTOCK: Probabilistic relational stock market predictions using diffusion models,” inIEEE Intl. Conf. Acoust., Speech and Signal Process. (ICASSP). 2024, pp. 7335–7339, IEEE

    work page 2024

  28. [28]

    DHMoE: Diffusion generated hierar- chical multi-granular expertise for stock prediction,

    Weijun Chen and Yanze Wang, “DHMoE: Diffusion generated hierar- chical multi-granular expertise for stock prediction,”AAAI Conf. Artif. Intell., vol. 39, no. 11, pp. 11490–11499, 2025

    work page 2025

  29. [29]

    Graph-aware diffusion for signal generation,

    Sergio Rozada, Vimal K B, Andrea Cavallo, Antonio G. Marques, Hadi Jamali-Rad, and Elvin Isufi, “Graph-aware diffusion for signal generation,” inIEEE Intl. Conf. Acoust., Speech and Signal Process. (ICASSP), 2026, pp. 461–465

    work page 2026

  30. [30]

    Diffusion model based resource allocation strategy in ultra-reliable wireless networked control systems,

    Amirhassan Babazadeh Darabi and Sinem Coleri, “Diffusion model based resource allocation strategy in ultra-reliable wireless networked control systems,”IEEE Commun. Lett., 2024

    work page 2024

  31. [31]

    Diffsg: A generative solver for network optimization with diffusion model,

    Ruihuai Liang, Bo Yang, Zhiwen Yu, Bin Guo, Xuelin Cao, M ´erouane Debbah, H. Vincent Poor, and Chau Yuen, “Diffsg: A generative solver for network optimization with diffusion model,”IEEE Commun. Mag., vol. 63, no. 6, pp. 16–24, 2025

    work page 2025

  32. [32]

    Generative diffusion models for resource allocation in wireless networks,

    Yi ˘git Berkay Uslu, Samar Hadou, Shirin Saeedi Bidokhti, and Alejandro Ribeiro, “Generative diffusion models for resource allocation in wireless networks,” inIEEE Intl. Wrksp. Computat. Advances Multi-Sensor Adaptive Process. (CAMSAP), 2025, pp. 201–205

    work page 2025

  33. [33]

    Diffu- sion model for multiple antenna communication,

    Jia Guo, Xiaoxia Xu, Yuanwei Liu, and Arumugam Nallanathan, “Diffu- sion model for multiple antenna communication,”IEEE Commun. Mag., vol. 63, no. 10, pp. 44–50, 2025

    work page 2025

  34. [34]

    Deterministic score-based diffusion model for channel estimation in ris-assisted mimo systems,

    Zhizhou He, Fabien H ´eliot, and Yi Ma, “Deterministic score-based diffusion model for channel estimation in ris-assisted mimo systems,” inIEEE Intl. Wrkshp. Signal Process. and Artif. Intell. for Wireless Commun. (SPAWC), 2025, pp. 1–5

    work page 2025

  35. [35]

    Gen- erating high dimensional user-specific wireless channels using diffusion models,

    Taekyun Lee, Juseong Park, Hyeji Kim, and Jeffrey G. Andrews, “Gen- erating high dimensional user-specific wireless channels using diffusion models,”IEEE Trans. Wireless Commun., vol. 25, pp. 2907–2921, 2026

    work page 2026

  36. [36]

    U-net: Convolu- tional networks for biomedical image segmentation,

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-net: Convolu- tional networks for biomedical image segmentation,” inMedical Image Comput. and Computer-Assisted Intervention (MICCAI). 2015, vol. 9351 ofLecture Notes in Computer Science, pp. 234–241, Springer

    work page 2015

  37. [37]

    nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,

    Fabian Isensee, Paul F. Jaeger, Simon A. A. Kohl, Jens Petersen, and Klaus H. Maier-Hein, “nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,”Nature Methods, vol. 18, no. 2, pp. 203–211, 2021

    work page 2021

  38. [38]

    Convolutional neural network architectures for signals sup- ported on graphs,

    Fernando Gama, Antonio G. Marques, Geert Leus, and Alejandro Ribeiro, “Convolutional neural network architectures for signals sup- ported on graphs,”IEEE Trans. Signal Process., vol. 67, no. 4, pp. 1034–1049, 2019

    work page 2019

  39. [39]

    Hierarchical graph representation learn- ing with differentiable pooling,

    Rex Ying, Jiaxuan You, Christopher Morris, Xiang Ren, William L. Hamilton, and Jure Leskovec, “Hierarchical graph representation learn- ing with differentiable pooling,” inAdvances in Neural Info. Process. Syst., 2018, vol. 31, pp. 4805–4815

    work page 2018

  40. [40]

    Graph u-nets,

    Hongyang Gao and Shuiwang Ji, “Graph u-nets,” inIntl. Conf. on Machine Learning (ICML), 2019, vol. 97 ofPMLR, pp. 2083–2092

    work page 2019

  41. [41]

    Striving for simplicity: The all convolutional net,

    Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller, “Striving for simplicity: The all convolutional net,” inIntl. Conf. on Learning Repr. (ICLR), Workshop Track, 2015

    work page 2015

  42. [42]

    Multi-scale context aggregation by dilated convolutions,

    Fisher Yu and Vladlen Koltun, “Multi-scale context aggregation by dilated convolutions,” inIntl. Conf. on Learning Repr. (ICLR), 2016

    work page 2016

  43. [43]

    Graph neural networks: Architectures, stability, and transferability,

    Luana Ruiz, Fernando Gama, and Alejandro Ribeiro, “Graph neural networks: Architectures, stability, and transferability,”Proc. IEEE, vol. 109, no. 5, pp. 660–682, 2021

    work page 2021

  44. [44]

    Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation

    Yoshua Bengio, Nicholas L ´eonard, and Aaron Courville, “Estimating or propagating gradients through stochastic neurons for conditional computation,”arXiv preprint arXiv:1308.3432, 2013

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  45. [45]

    Reversible instance normalization for accurate time- series forecasting against distribution shift,

    Taesung Kim, Jinhee Kim, Yunwon Tae, Cheonbok Park, Jang-Ho Choi, and Jaegul Choo, “Reversible instance normalization for accurate time- series forecasting against distribution shift,” inIntl. Conf. on Learning Repr. (ICLR), 2022

    work page 2022

  46. [46]

    Stochastic beams and where to find them: The gumbel-top-k trick for sampling sequences without replacement,

    Wouter Kool, Herke Van Hoof, and Max Welling, “Stochastic beams and where to find them: The gumbel-top-k trick for sampling sequences without replacement,” inIntl. Conf. on Machine Learning (ICML). PMLR, 2019, pp. 3499–3508. 13 APPENDIXA ACCELERATEDDDIM SAMPLING The base DDIM sampler (9) deferred the specification of its noise scaleσ k. We provide it here...

    work page 2019