Directional diffusion models for graph representation learning

In recent years, diffusion models have achieved remarkable success in various domains of artificial intelligence, such as image synthesis, super-resolution, and 3D molecule generation. However, the application of diffusion models in graph learning has received relatively little attention. In this paper, we address this gap by investigating the use of diffusion models for unsupervised graph representation learning. We begin by identifying the anisotropic structures of graphs and a crucial limitation of the vanilla forward diffusion process in learning anisotropic structures. This process relies on continuously adding an isotropic Gaussian noise to the data, which may convert the anisotropic signals to noise too quickly. This rapid conversion hampers the training of denoising neural networks and impedes the acquisition of semantically meaningful representations in the reverse process. To address this challenge, we propose a new class of models called {\it directional diffusion models}. These models incorporate data-dependent, anisotropic, and directional noises in the forward diffusion process. To assess the efficacy of our proposed models, we conduct extensive experiments on 12 publicly available datasets, focusing on two distinct graph representation learning tasks. The experimental results demonstrate the superiority of our models over state-of-the-art baselines, indicating their effectiveness in capturing meaningful graph representations. Our studies not only provide valuable insights into the forward process of diffusion models but also highlight the wide-ranging potential of these models for various graph-related tasks.

Paper

References (35)

Scroll for more · 23 remaining

Similar papers

Peer review

Reviewer ejpR7/10 · confidence 2/52023-06-28

Summary

The authors present directional diffusion model (DDM) to learn graph / node representation. Compared to vanilla diffusion based representation learning techniques, DDM adds a batch based mean and variance, as well as preserving direction of diffusion. The authors then demonstrate that their model perform better than other models on representating learning by experimenting on various dataset and compare with other methods.

Strengths

- It is a nice finding by the authors that compared to diffusion model for generative modeling, diffusion model for representation learning does not require sampling from final distribution, thus we don't need to know the limiting behavior of the diffusion process. - The experiments supports the authors' claim on the performance of the model.

Weaknesses

- The authors should include a brief introduction on how diffusion models work on representation learning

Questions

- A lot of citations by the paper, including some baselines like Infograph and GraphMAE, come from arxiv. Are these methods reliable? - Directional diffusion preserves direction, so in recovery process, won't the representation fail to learn the direction of the entries? If it doesn't learn the directions, does directions don't matter?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

2 fair

Contribution

4 excellent

Limitations

N/A

Reviewer 8dwQ6/10 · confidence 4/52023-07-03

Summary

This paper offers a good study on anisotropic and directional structure learning of the diffusion model. The authors first conduct a statistical analysis emphasizing the importance of anisotropic structure in graph dataset and demonstrate the disadvantage of the vanilla diffusion model through signal-to-noise ratio. Later on, they propose their work, which is a new pipeline considering preserving the characteristic of data in the diffusion process. Its idea is simple: adding directional diffusion on data, preserving more information than vanilla diffusion. Its model is simple, leveraging 4 layers GNNs and one layer MLP, it achieves improvement in several benchmarks. The idea is elegant.

Strengths

Good idea! The work not only offers inspiration to the graph learning community, but it also contributes to the diffusion model. The two constraints added in the work are elegant and proper.

Weaknesses

The writing is not so well, especially the formula part. There are too many mistakes there. Their performances aren't satisfying, but I believe the underlying problem doesn't come from the pipeline, it comes from the simple model. There aren't convergence proofs for their learning architecture. Noticed that the diffusion model, which stems from the quasi-static process, possesses a good property of converging. I hope to see proof of the effectiveness of the constraints adding to the original process.

Questions

1. Do you have proof of the effectiveness of your idea? 2. Can you offer more experimental results?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

4 excellent

Limitations

1. They haven't provided limitations for their work. 2. Though the idea is inspirational, the writing is not good, and there are too many writing mistake.

Reviewer tkhv5/10 · confidence 5/52023-07-06

Summary

This paper presents a method named Directional Diffusion Model (DDM) for unsupervised representation learning, targeting applications in graph and node classification. The model's performance is evaluated on various benchmark datasets and compared to both unsupervised and supervised models. The results demonstrate that DDM can effectively learn meaningful graph- or node-level representations. Furthermore, the paper offers some exploration of how different types of noise impact the learning process.

Strengths

1. The authors present empirical evaluations of DDM on multiple benchmark datasets, which validate the effectiveness of the proposed method to some extent. 2. The paper includes an investigation into the impact of different types of noise, which adds depth to the analysis and understanding of the proposed method.

Weaknesses

1. The SVD visualizations could be more insightful. It's observed that the 2D projections computed from graph datasets appear biased and are predominantly on the right plane. However, the methodology behind these visualizations remains unclear. Are they based on singular values? Further explanation would be beneficial. 2. The Signal-to-Noise Ratio (SNR) plots in the supplementary material show comparable results under white noise and directional noise. This implies that the DDM may not offer significant advantages on these datasets. The authors should provide some clarification. 3. The paper is primarily empirical and lacks theoretical foundations, which might limit its generalizability. 4. The algorithm snippet doesn't seem to provide sufficient implementation details, which might impede attempts to reproduce the study.

Questions

1. Please clarify the issues raised in the weaknesses section. 2. How does DDM compare with widely-known baselines in terms of performance, and are there any specific scenarios where DDM particularly excels or falls short? 3. Regarding the means and deviations at line 170: are these calculated relative to the node features at time 0, or do they consider other time steps? In its current form, I believe the paper does not meet the rigorous quality and contribution standards expected at a top-tier conference like NeurIPS.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

The authors did not explicitly state any limitations of their study. However, the identified weaknesses could be viewed as potential limitations. These include the lack of theoretical foundations and potential issues in reproducibility due to insufficient implementation details. Additionally, the paper could benefit from a clear discussion of the limitations of DDM in handling specific scenarios or types of data.

Reviewer jRHd5/10 · confidence 3/52023-07-07

Summary

This study proposed adding directional noise on learning anisotropic graphs using diffusion models. The study adds new perspectives in exploring the anisotropic structures in graph data. The numerical results are promising to support the authors' ideas.

Strengths

I find this is an interesting paper - the rational is convincing, and the authors performed extensive experiments to support the idea. The discussions on the noise and ablation studies provide further insights into understanding the usefulness of adding directional noises. The paper provides a valuable perspective in understanding diffusion models for graph learning.

Weaknesses

The authors mainly prove the utility of the proposed approach through experiments; there is a lack of theorectical proof.

Questions

The authors have performed extensive experiments, but it would add more evidence if the authors can add some theoretical proof. For figure 3, does each dot represent a graph? Are they from simulated datasets?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors have discussed several future directions that can be used to improve the current model.

Reviewer BUxE6/10 · confidence 3/52023-07-24

Summary

This work proposes a class of diffusion models to improve the accuracy of graph representation learning. The model incorporates both data-dependent and anisotropic noise in the forward noising process, by scaling its magnitude and direction based on each coordinate of the data. This structured noise maintains the signal present in the data over longer time windows than using standard isotropic white noise during the forward noising process. The authors show that this model improves upon state-of-the-art methods on a large collection of benchmark datasets. Moreover, they perform an ablation study to understand the effect of the two proposed modification to the noise process.

Strengths

- The authors visually explain and empirically demonstrate the effect of using non-isotropic noise in the forward diffusion process to improve classification tasks. These noise processes provide clear intuition why this is preferred to white noise for these tasks. - The authors propose a novel strategy for extracting graph representations based on time-dependent denoiser that combines graph neural networks and UNet architectures. - The application of diffusion models to graph representation learning is novel and the new model is shown to yield superior results to existing algorithms.

Weaknesses

It would be great for the authors to comment and compare with other non-isotropic noise processes that have been proposed for diffusion models for sampling and generative modeling. Some examples outside of the graph representation learning context are: * Score-based Denoising Diffusion with Non-Isotropic Gaussian Noise Models, Vikram Voleti, Christopher Pal, Adam Oberman, 2022 * Blurring Diffusion Models, Emiel Hoogeboom, Tim Salimans, 2023 In addition to the changes in the noise process, the authors propose a specific architecture for the denoiser, and selection of representative features from their hidden layers. This architecture could be relevant on its own to extract representations of the dataset, without the denoising time components. Do any of the compared methods investigate how this denoiser architecture, without the diffusion model, would perform for representation learning. This might be helpful as an additional ablation study to see the effect of the architecture. The metrics used to evaluate their experiments can be described in more detail. The values for the results in Tables 1,2,3 were not clearly mentioned in the caption or the main text. Mathematical equations may also be helpful to precisely describe the accuracy measurements in Figure 5. The authors comment that the word vectors often exhibit greater anisotropy, which yields superior performance in node classification tasks. It would be great if the authors could quantify this to validate this claim that the performance improves with more anisotropy.

Questions

An important choice for defining the representations are the user selected time-steps where the outputs are extracted from the denoising network. Do the authors have a guideline or recommendation (e.g., based on the empirical studies) for how to choose these time steps. This seems particularly relevant given that the SNR does not degrade monotonically for some datasets (e.g., IMDB in Figure 2). Even if the SNR in standard diffusion models decay quickly with increasing time, how do the results presented here compare to a baseline diffusion model with isotropic noise where the time-steps are chosen "optimally". For example, what if the time-steps are chosen all near the initial time, $t = 0$, when there is still signal contained in the noisy data?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

The authors show improved results on almost all datasets. It would be great if the authors could comment on what leads to the similar performance on the Ogbn-arxiv dataset. Is it because the data is more isotropic in this case? When do the authors expect the algorithm to under-perform for the evaluated tasks? In addition to the questions above, it would be good to address these minor comments: - Define the acronym for their proposed framework, DDM - Include standard errors for the results in Table 3 - Explain why $f_\theta$ depends on $A$ in equation (4). It is sufficient to highlight that this is the structure of a GNN, which may be helpful for some readers. - Clarify in Figure 4 that the directional noise is only added to $X_0$ and not $A$.

Reviewer emkS7/10 · confidence 4/52023-07-25

Summary

The authors address the gap in unsupervised graph representation learning by exploring the use of diffusion models. They propose directional diffusion models that incorporate data-dependent, anisotropic, and directional noises in the forward diffusion process to better handle anisotropic structures in graphs. Experiments on publicly available datasets showcase the superiority of their models over state-of-the-art baselines, demonstrating their effectiveness in capturing meaningful graph representations. Overall, the paper presents a compelling approach that contributes to the advancement of unsupervised graph representation learning.

Strengths

1. **Motivation** The introduction is well-motivated, providing a thorough explanation of the challenge and task at hand. The authors go beyond textual descriptions and use simple visualizations on both real and synthetic data to demonstrate their points effectively. This approach enhances the clarity and understanding of the presented research, making it accessible to a wider audience.   2. **Method** The authors present a straightforward yet effective solution for incorporating directional noise into node embeddings. This approach effectively addresses the challenge posed by anisotropic structures in various domains, including graphs and potentially text data. Their proposed method demonstrates promising results in handling directional noise and enhancing the quality of node and graph embeddings.   3. **Architecture** I appreciate the authors intention to adapt the well-known and effective U-Net architecture from the image domain to the graph domain. The incorporation of skip connections in the U-Net is particularly relevant for denoising tasks. This thoughtful adaptation enhances the model's ability to handle graph-related denoising effectively.   4. **Experiments** The authors conduct a comprehensive comparison with numerous baselines across multiple datasets. Furthermore, their evaluation settings, which involves 10-fold cross-validation with standard deviation after five runs, are robust and reliable. This rigorous evaluation methodology ensures the validity and statistical significance of their results.

Weaknesses

1. **missing releted work** There are existing works in the intersection of graphs and diffusions are missing, contradicting the authors statement "To the best of our knowledge, there have been no works for diffusion-model-based graph representation learning.". Some for example: - Niu, Chenhao, et al. "Permutation invariant graph generation via score-based generative modeling." International Conference on Artificial Intelligence and Statistics. PMLR, 2020. - Xu, Minkai, et al. "Geodiff: A geometric diffusion model for molecular conformation generation." International Conference on Learning Representations, 2022. - Vignac, Clement, et al. "Digress: Discrete denoising diffusion for graph generation." International Conference on Learning Representations, 2023. Furthermore, some simple techniques can be applied to create a smoother SNR curves over the different diffusion steps. For instance: - Chen, Ting. "On the importance of noise scheduling for diffusion models." arXiv preprint arXiv:2301.10972 (2023). The same problem was presented in it over the image domain (on high-resolution images), and the solution was to use different noise schedulers. Why not simply try this trick?   2. **missing details** There are missing details in the paper that makes it hard to fully understand and reproduce the papers results. For example: - "µ and σ are calculated using graphs within the batch.” -- what is done during inference? it is EMA over what was been seen in training time? does it handle only batch inference? - They paper did not state how exactly the entire graph-level representation is obtained, it is sum/mean of all node representations? - Algorithm 2 in appendix, line 6: what is exactly $A_t$? should it be just $A$?

Questions

1. "µ and σ are calculated using graphs within the batch.” -- what is done during inference? why not just pre-calculate those on the entire train split? 2. "allowing the latent code to contain meaningful compressed knowledge while preventing the decoder from degrading into an identical mapping” -- why will that happen that requires a dedicate MLP? what is this different from the simple U-Net used in the image domain? 3. Algorithm 2 in appendix, line 6, what is $A_t$? 4. In some cases, the SNR goes even higher for larger diffusion steps (like in the Computer dataset for instance), why is that? 5. How is the entire graph-level representation obtained? 6. Referring to figure 2, shouldn't the SNR be 0 in the last diffusion steps? It doesn't reach zero as the directional noise keeps them separated, is that a wanted outcome? If so, why? 7. Why not apply just a different noise scheduler? instead of linear, coisne for instance. Similar observation was shown on high-resolution images and different schedulers where proposed [1]. I know it is still isotropic noise, so at least show results as a baseline/ablation. ___ [1] Chen, Ting. "On the importance of noise scheduling for diffusion models." arXiv preprint arXiv:2301.10972 (2023).

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The authors address the limitations of their work, and support their claims with experiments and ablations.

Reviewer tkhv2023-08-11

thank you for the responses

The feedback from the reviewers has addressed some of my concerns, prompting me to increase the score to 5.

Reviewer 8dwQ2023-08-12

Response to the authors' rebuttal.

Thanks for offering kind responses to my questions. I hope you could correct writing mistakes and make the paper more readable. Please offer a elegant theorectical proof as soon as possible, the idea of the paper is innovative and good.

Reviewer jRHd2023-08-15

I appreciate the authors' responses. I don't have any other questions.

Reviewer emkS2023-08-18

I appreciate the response from the authors which answered almost all my concerns. The only thing I'm not yet convinced about is why the SNR doesn't reach zero - isn't a graph, for example, with all possible edges have SNR=0? I'd appriciate an explanation. Anyway, I'm upgrading my score to accept, as besides this issue, all my questions where fully answered.

Authorsrebuttal2023-08-20

Thank you very much for your active engagement. In our SNR analysis, we use the LDA method to quantify the signal-to-noise ratio of the **node features** for the node or graph class labeling (this design aims to be consistent with downstream tasks ). Since our batch-dependent noise is always correlated with $x$, the $x_{t}$ will always be more Informative than pure white noise, so it is still possible that the SNR doesn't reach zero in the last diffusion steps. If you have any remaining concerns regarding our article, we cordially welcome you to present them, and we shall exert our utmost efforts to address them comprehensively.

Reviewer BUxE2023-08-19

We thank the authors for their detailed response! The feedback from the reviewers has addressed some of my concerns, so I've raised my score. I would still appreciate some comparisons with non-isotropic noise or with choosing different schedules for the noise as baselines for the proposed data-dependent noise (in the absence of theoretical results on the method).

Authorsrebuttal2023-08-20

Thank you very much for your active engagement. We are willing to provide more comparisons with non-isotropic noise or with choosing different schedules here ( Same as what we provided to Reviewer 1). | Noise schedule function | Noise type | Citeseer | PubMed | MUTAG | |--------------------------|-------------|----------|--------|-------| | cosine (s=0,e=1,τ = 1) | DDM | 0.715 | 0.824 | 0.867 | | cosine (s=0,e=1,τ = 1) | White Noise | 0.371 | 0.453 | 0.692 | | sigmoid (s=-3,e=3,τ = 1) | DDM | 0.735 | 0.803 | 0.889 | | sigmoid (s=-3,e=3,τ = 1) | White Noise | 0.672 | 0.606 | 0.689 | | sigmoid (s=0, e=3,τ = 1) | DDM | 0.710 | 0.806 | 0.877 | | sigmoid (s=0, e=3,τ = 1) | White Noise | 0.581 | 0.434 | 0.691 | It shows that. Different schedulers do indeed influence the final effectiveness of the model. However, due to the data-independent nature of white noise, its ultimate performance heavily relies on the hyperparameters of the scheduler. Yet, irrespective of the scheduler employed, our proposed data-dependent anisotropic noise consistently yields superior outcomes. This confirms our impression of anisotropic structures in section 1. We will add this experiment to the paper in the subsequent version.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC