acceptodds
Under review as a conference paper at ICLR 2027

Multiview Contrastive for Graph Diffusion

Abstract

Graph diffusion models have recently emerged as powerful generative models for structured data. They learn data distributions by gradually transforming noise into samples through a reverse diffusion process, where the training objective can be interpreted as minimizing the discrepancy between generated and target distributions. From an information-theoretic perspective, we revisit the distribution matching objective and reveal that the entropy term is an intrinsic component of this objective. This observation indicates that preserving the information content carried by the data distribution can provide a complementary perspective for improving generative modeling. Furthermore, we show that maximizing mutual information between multiple views provides a tractable surrogate for entropy maximization. Based on this insight, we propose a multiview contrastive graph diffusion learning (MCGD) method, a framework that introduces entropy-oriented contrastive regularization into graph diffusion models. MCGD constructs stochastic views from denoised graph predictions and performs hierarchical contrastive regularization from two complementary perspectives: Graph-level contrastive learning, which enforces consistency in the prediction space, and Semantic-level contrastive learning, which aligns high-level graph representations in a latent semantic space. By jointly preserving structural information and high-order graph semantics, MCGD enables diffusion models to better capture complex graph distributions. Extensive experiments on general and molecular graph generation benchmarks demonstrate that MCGD improves generation quality and structural fidelity consistently.

Primary area:
generative models
Keywords:
Graph Diffusion Model, Multiview learning, Information theory, Contrastive learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_CSCDOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MCGD, whose core idea is to introduce multiview contrastive learning into the standard denoising training of graph diffusion models. By constructing two augmented views from the same predicted graph and enforcing consistency at both the graph-structure level and the high-level semantic level. Experimental results show that this method improves generation quality on both general graph generation and molecular graph generation tasks, demonstrating the effectiveness of the proposed mtehod.

Critical strengths
  • Introducing auxiliary representation-level constraints into diffusion training is a reasonable idea.
  • MCGD achieves strong results across several evaluation metrics.
Critical weaknesses
  • Theoretical Analysis. In Eq. (3), H(x)H(x) is the entropy of the data distribution, so it is constant with respect to θ\theta and provides no gradient signal. Eq. (4) only maximizes a lower bound on this constant, which does not show that pθp_\theta becomes closer to qq. If the goal is to increase the entropy of the generated distribution, the authors should explain why this would improve fidelity rather than harm it.
  • Methods. The edge encoder applies an MLP to a flattened, zero-padded adjacency matrix, making the representation dependent on node ordering. This weakens the claim that it captures high-order graph semantics. Its input size also scales as O(n2)O(n^2).
  • Experiments. Prior work typically includes more datasets, such as Community-small, Grid, and MOSES. The paper should report additional results on these benchmarks. The ablation results do not fully support the conclusions. On QM9, adding the regularizers reduces validity (99.79 → 99.61). Both single-loss variants also outperform full MCGD on scaffold similarity.
Questions for authors

Please see the Weaknesses.

Reviewer_7qyqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MCGD (Multiview Contrastive Graph Diffusion learning), which regularizes graph diffusion models with additional contrastive losses. They show that entropy plays an important role in identifying the content of the target distribution. Adding to this, they propose contrastive learning objectives to the denoising objective on augmented graph views. They use two kinds of contrastive loss, where the first focuses on graph-level contrastive loss, and the second on semantic-level contrastive loss. By adding this to the diffusion loss, they show that their method performs superior to different graph generation baselines.

Critical strengths

Here are some strengths of the paper in my opinion:

  1. They frame their work by showing that the information theoretic analysis of graph diffusion which encourages the generated distribution to preserve information beyond likelihood matching

  2. They provide two useful surrogates to provide the additional information to the generation process

  3. They provide ablation studies on how the two contrastive losses affect the performance

Critical weaknesses

Here are some weaknesses of the paper in my opinion:

  1. The theory presented for using contrastive learning doesn't seem strong. The entropy term has no effect on optimizing the diffusion model. Its appearance in the KL decomposition does not establish that an additional entropy-maximization objective is necessary.

  2. The improvements seem very small and it seems like the model is performing well on everything. What are some weaknesses of the model? Or is it because the datasets are easy?

  3. The datasets used seem small scale. Can the authors provide results on large-scale datasets like MOSES or others?

Questions for authors

Please look at the weaknesses section

1 comment
Reviewer_7qyqOpenReview ↗

Reply to Authors

Dear Authors,

Thank you for your rebuttal. I will keep my current score because the reply still doesn't address some of the concerns raised. First, the authors are still working on bigger datasets, which is not currently ready. Second, my second weakness point is still there where the improvements are very small. Because of that, I will keep my current score.

Reviewer_dDUHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper augments graph diffusion training with graph-level consistency and semantic-level InfoNCE objectives. The method constructs two stochastic views from denoised graph predictions, aligns their predicted structures, and contrasts learned graph representations; a training-only semantic encoder is discarded at sampling time. The motivation is plausible, and the engineering is compatible with existing pipelines, but the theoretical claims outrun the derivation, the graph-level loss raises a collapse concern that is not resolved, and the empirical gains are uneven and reported without uncertainty.

Critical strengths
  • This paper is easy to follow.

  • The method addresses a real limitation of denoising-only training—that generated samples can satisfy local reconstruction objectives without preserving global or semantic structure.

  • The semantic encoder is training-only and discarded at sampling, so the addition requires no new generation procedure and can in principle drop into existing graph diffusion pipelines.

Critical weaknesses
  • The novelty and technical contribution of this paper are limited. This paper combines some existing machine learning and graph learning methods together.

  • The experiments section can be further improved. For example, the improvements are not consistent across datasets and metrics, and ablations do not show each loss helps uniformly. Planar statistics are better without the graph-level term, and clustering gains on Enzymes are smaller than other baselines.

  • The graph-level objective is derived by dropping the negative-pair contribution on the assumption that its gradient is negligible, after which it reduces to a pairwise squared-error consistency loss. That loss is not a contrastive or mutual-information objective and can favor collapsed predictions; the paper does not establish why the denoising term reliably prevents collapse.

Questions for authors

see the weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.