Discrete-state Continuous-time Diffusion for Graph Generation

Graph is a prevalent discrete data structure, whose generation has wide applications such as drug discovery and circuit design. Diffusion generative models, as an emerging research focus, have been applied to graph generation tasks. Overall, according to the space of states and time steps, diffusion generative models can be categorized into discrete-/continuous-state discrete-/continuous-time fashions. In this paper, we formulate the graph diffusion generation in a discrete-state continuous-time setting, which has never been studied in previous graph diffusion models. The rationale of such a formulation is to preserve the discrete nature of graph-structured data and meanwhile provide flexible sampling trade-offs between sample quality and efficiency. Analysis shows that our training objective is closely related to generation quality, and our proposed generation framework enjoys ideal invariant/equivariant properties concerning the permutation of node ordering. Our proposed model shows competitive empirical performance against state-of-the-art graph generation solutions on various benchmarks and, at the same time, can flexibly trade off the generation quality and efficiency in the sampling phase.

Paper

References (87)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 9BTR6/10 · confidence 4/52024-06-23

Summary

This paper is an extension of current graph diffusion generation to discrete-state and continuous-time version, and proposes DisCo model. The authors also prove permutation-equivariance/invariance of the proposed diffusion framework, sampling and training loss.

Strengths

1. Continuous-time diffusion is a generlization of different strategies, e.g. DDPM and NCSN, so it is more powerful and flexible. Discrete state is also suitable for graph data becaused of discrete edge connection and node classes. This paper integrates these two advantages into a framework, and its good performance is expected. 2. The wrting quality and organization is good and easy to follow.

Weaknesses

The selected baselines are not enough. Only DiGress is compared as graph diffusion generation, but other generation methods based on diffusion have been developed. I only list two relevant baselines [1-2], but I convince that there are more relevent baselines. [1] Chen, X., He, J., Han, X., & Liu, L. P. (2023). Efficient and degree-guided graph generation via discrete diffusion modeling. arXiv preprint arXiv:2305.04111. [2] Fu, X., Gao, Y., Wei, Y., Sun, Q., Peng, H., Li, J., & Li, X. (2024). Hyperbolic Geometric Latent Diffusion Model for Graph Generation. arXiv preprint arXiv:2405.03188.

Questions

As shown in table 3, DisCo-MPNN is much more efficient than DisCo-GT. But from eq. (7) and (8), DisCo-MPNN also conducts aggregation over the whole graph, so its complexity is the same as graph transformer. How to explain this point?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

I recommend the authors supplement more relevant baselines to enrich the experiments.

Reviewer XoYX5/10 · confidence 3/52024-07-07

Summary

The paper presents DISCO (Discrete-State Continuous-Time Diffusion), a novel framework for graph generation. DISCO formulates the graph diffusion process in a discrete-state continuous-time setting, preserving the discrete nature of graph data while allowing for flexible sampling trade-offs between quality and efficiency. The model integrates concepts from continuous-time Markov Chains (CTMC) and employs a graph-to-graph neural network backbone for efficient graph generation.

Strengths

- the paper is well written. most sections are easy to follow - the proposed research question is important (efficiency is indeed a limitation of current discrete graph generation model) - the proposed method is technically sound

Weaknesses

> w1: the proposed seems to be a rather straightforward derivation from the existing tools (replacing discrete markov chain with continuous time) > w2: no implementation available > w3: there is a disconnection between the proposed motivation (sampling motivation) and the (theoretical) discussion of the proposed method (see questions below) > w4: most theoretical results seem to be a standard derivation from existing literature

Questions

> Q1: the motivation for making the original discrete transition process continuous is to improve the sampling efficiency. Has this been achieved in the proposed method? If so, what is the improvement and where exactly does this improvement come from?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

N.A

Reviewer ohXc5/10 · confidence 3/52024-07-09

Summary

This paper categorizes existing graph diffusion models into 4 types according to whether the space of states and time steps are discrete or continuous. The authors design DISCO, the first discrete-state continuous-time graph diffusion generative model. By incorporating continuous-time Markov chains while preserving the discrete nature of graph data, DISCO enables more flexible sampling and decouples it from model’s training. Furthermore, this paper explores a novel and concise message-passing neural network as a replacement for complex graph transformer, while maintaining competitive performance.

Strengths

1. The motivation of this work is clear and easy to understand. 2. This paper provides comprehensive mathematical proofs to help to clearly understand how to apply continuous-time Markov chains to graph diffusion generative models

Weaknesses

1. Some symbols in the formulas of this paper are not clear in the context, which reduces readability 2. This paper seems to merely combine continuous-time Markov chains with discrete graph data, essentially integrating two previous works, and the innovation is slightly insufficient.

Questions

Besides more flexible sampling, do graph diffusion models based on continuous time have any other advantages compared to discrete-time models? Does the flexibility of sampling provide any particular benefits to generating new data?

Rating

5

Confidence

3

Soundness

3

Presentation

2

Contribution

2

Limitations

Yes, the authors addressed limitations, potential negative societal impact, and mitigation.

Reviewer Tpg46/10 · confidence 3/52024-07-12

Summary

The paper introduces a novel graph generation method based on the continuous-time Monte Carlo Markov Chain by adapting the previous work [7]. The proposed architecture comes in two variants: one utilizing a graph Transformer and another using a regular message-passing neural network. The performance of the model is evaluated on artificial planar and SBM networks, as well as on real datasets for the molecule graph generation task.

Strengths

- The paper proposes a novel graph generation method relying on a continuous-time Markov Chain-based diffusion generative model. - The proposed architecture is supported by theoretical analysis. - The model is evaluated on both artificial and real datasets and shows comparable performance with respect to the state-of-the-art methods.

Weaknesses

- In the real datasets, for the molecule graph generation task, baseline methods such as GraphINVENT and DiGress show comparable or better performance in some metrics. - The paper relies on adapting a previous continuous-time Markov Chain-based diffusion generative model for the graph domain. Thus, it doesn’t present substantial new material in terms of novelty.

Questions

**Questions** - The same metrics are not used across various tables (Tables 2 and 6-8), and some tables lack error terms. This inconsistency makes it difficult to compare results comprehensively. Can the authors update the results to enhance the clarity and reliability of the results? - Can the proposed architecture be applicable to generate weighted and directed networks? - The related work section is very brief, possibly due to page limitations but an extended version can be added in the appendix. **Typos and Errors** - In Lines 94-95, the set membership symbol should be the subset symbol. - In Lines 193-197, it is assumed that J_{f_i, f^bar_i} follows Poisson distribution so how can the conditions \sum_{ f^bar_i} J_{f_i, f^bar_i} = 1 and J_{f_i, s} > 1 hold for some s because all entries are non-negative?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors addressed the limitations of the model in the appendix.

Authorsrebuttal2024-08-11

We sincerely appreciate your thoughtful review and are glad to have the opportunity to clarify your questions. Thank you for your recognition of our work!

Reviewer Tpg42024-08-09

Thanks for your response

I appreciate the authors' efforts in addressing the points I’ve indicated, and I increased my score.

Reviewer XoYX2024-08-10

Thank you for the diligent response. However, I still found my main concerns not fully addressed. As the main motivation for using a continuous Markov model instead of a discrete one is to improve the sampling efficiency, one should expect to see more evaluation and evidence to support and validate that this has been achieved. In addition, the result in Table 9 shows that the proposed method is robust to the sampling step and as mentioned in the paper, DIGRESS also has this property. This does not show there is an improvement in sampling efficiency. Q: Why can't $\tau$-leaping be used in the discrete model? Can it be applied in other continuous graph generation such as GDSS?

Authorsrebuttal2024-08-11

Further Discussion

Dear Reviewer XoYX, We appreciate your insightful response and are glad to clarify your concerns. We sincerely invite the reviewer to have further discussions. > **Q1.** Why can't $\tau$-leaping be used in the discrete model? Can it be applied in other continuous graph generation such as GDSS? **A1.** We clarify this as follows. 1. **$\tau$-leaping [1-3] is designed for continuous-time discrete-state** stochastic processes (CTMC). It **cannot** be applied to discrete-time models (e.g., Digress [4]) because $\tau$-leaping needs the rate matrices as input, but **discrete-time models do not have rate matrices**. 2. It **cannot** be applied to GDSS because GDSS is a **continuous-state** model, which is not characterized by the rate matrices. > **Q2.** The result in Table 9 shows that the proposed method is robust to the sampling step and as mentioned in the paper, DIGRESS also has this property. This does not show there is an improvement in sampling efficiency. **A2.** We would like to answer the reviewer' question from the following perspectives: 1. Theoretically, Digress cannot have such sampling flexibility. For example, **if the model is trained with 500 forward steps, technically, its reverse process has to include 500 steps**. In other words, for a Digress model trained with 500 forward steps, doing the reverse process in 100 steps is hard to ensure that the generated samples are from the desired generation distribution. 2. Empirically, we have conducted the following experiments to answer the reviewer's question better. We train a Digress model in **500 forward steps** and use it to **do reverse sampling in 500 steps**. We artificially freeze the reverse transition probability in every $\mathrm{round}(500/k)$ steps, where $k$ is the number of inferences. In this way, we have a fair comparison to show the sampling robustness regarding the number of model inferences in Table 1. The metric is validity, and the results of DisCo are shown in Table 4 of our submitted paper. Both DisCo and Digress use the graph Transformer as the backbone and the marginal distribution as the reference distribution. **DisCo is consistently more robust regarding the decrease in the number of inferences. In other words, DisCo has a better quality-efficiency tradeoff.** | \# of inferences | DisCo | Digress | |---|---|---| | 500 | 99.3$\pm$0.6 | 98.9$\pm$0.4 | | 100 | 98.7$\pm$0.5 | 95.0$\pm$1.1 | | 30 | 97.9$\pm$1.2 | 92.1$\pm$0.7 | | 10 | 95.3$\pm$1.9 | 87.9$\pm$0.8 | | 5 | 93.0$\pm$1.7 | 80.1$\pm$1.4 | | 1 | 76.1$\pm$2.3 | 67.1$\pm$2.9 | Table 1. Quality-efficiency tradeoff comparison on QM9 3. In addition, we did not find results from the Digress paper [4] showing that Digress is robust to the step size. If the reviewer could kindly give us a point on related results/statements, we will be happy to check them out. > **Q3.** As the main motivation for using a continuous Markov model instead of a discrete one is to improve the sampling efficiency, one should expect to see more evaluation and evidence to support and validate that this has been achieved. **A3.** As we mentioned in the submitted paper, Tables 4 and 9 show that the model's performance is robust when the sampling steps are decreased, achieving a decent generation quality-efficiency tradeoff. Lines 844-846 in the appendix detail the sampling steps used to search for the best results, much smaller than Digress's number of sampling steps. In addition, the results from Table 1 in this response should be supplementary to show the generation efficiency of DisCo, which will be included in the revised paper. ``` Reference: [1] Campbell, Andrew, et al. "A continuous time framework for discrete denoising models." NeurIPS 2022. [2] Gillespie, Daniel T. "Approximate accelerated stochastic simulation of chemically reacting systems." The Journal of chemical physics 115.4 (2001): 1716-1733. [3] Wilkinson, Darren J. Stochastic modelling for systems biology. Chapman and Hall/CRC, 2018 [4] Vignac, Clement, et al. "DiGress: Discrete Denoising diffusion for graph generation." ICLR 2023. ```

Reviewer XoYX2024-08-14

I thank the authors for their diligent responses. Most of my concerns and questions are addressed and I have raised my score accordingly. I will suggest the authors to make further emphasis on the sampling aspect in both the technical and experiment section for revised version, as that is the main the motivation of the paper.

Authorsrebuttal2024-08-14

We sincerely thank the reviewer for insightful questions, suggestions, and discussion. We appreciate your recognition of our work and will revise the paper according to the valuable feedback provided!

Reviewer 9BTR2024-08-10

Response to the rebuttal

Thank the authors for more experiments, and further explanation. I choose to keep my positive score.

Authorsrebuttal2024-08-11

We sincerely appreciate your thoughtful review and are glad to have the opportunity to clarify your questions. Thank you for your recognition of our work!

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC