- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes Manifold Alignment Learning (MAL), an auxiliary regularization objective for graph diffusion models. The authors argue that conventional sample-wise denoising objectives may not adequately preserve the global distributional structure of graph data. To address this, MAL estimates empirical score fields from minibatches of real graphs and denoised predictions using kernel density estimation (KDE), with kernel similarities computed through a pretrained graph encoder. The resulting score-field discrepancy is incorporated into the original diffusion training objective without modifying the backbone architecture or sampling procedure. The method is evaluated on four graph generative backbones (GDSS, GruM, DiGress, and DeFoG) and six molecular and general graph benchmarks. The experiments report improvements in several generation-quality metrics, including molecular validity, FCD, and graph structural statistics. The paper also provides hyperparameter sensitivity analyses and additional multi-seed results on QM9.
Critical strengths
-
Simple and broadly applicable regularization framework.
A notable strength is the modularity of the proposed approach. MAL can be integrated into different graph generative frameworks without modifying their underlying architectures or sampling procedures. The experiments across both continuous and discrete graph generation models support its practical applicability. This flexibility makes the approach potentially useful as an auxiliary training objective.
-
Encouraging empirical results across multiple benchmarks.
The paper reports consistent improvements on several generation-quality metrics across multiple backbones and datasets. In particular, the improvements in FCD and molecular validity on QM9 and ZINC250k are encouraging. The additional evaluations on general graph benchmarks further demonstrate the potential applicability of the approach beyond molecular generation.
Critical weaknesses
-
The Gaussian-kernel derivation does not cover the implemented field. Appendix A.4 derives the score identity using Gaussian weights and displacements in the same coordinate space: the raw displacement follows from differentiating the squared Euclidean distance in the kernel exponent. Section 3.3 then substitutes encoder-space weights, while Eq. 9 retains raw-coordinate displacements. Differentiating the encoder-space KDE with respect to the raw input instead introduces the encoder Jacobian and feature-space displacements, which are absent from Eq. 9. The same centroid formula therefore does not generally remain a score after this substitution. A justification is needed for the encoder-based field, since its score interpretation provides the mathematical basis for the paper's distribution-alignment claim.
-
The link from prediction alignment to correcting reverse sampling needs justification. Section 3.1 attributes sampling distortion to using a single conditional-mean prediction, while Eq. 9 aims to align the marginal distribution of predictions across a batch with the data distribution. A reverse transition depends on how each clean prediction relates to the current noisy state; matching prediction marginals alone does not determine that relationship. Appendix A.3 bounds discrepancy after applying a shared forward-corruption kernel to clean outputs and explicitly notes that it does not directly characterise reverse sampling. It therefore does not establish the claimed correction of the reverse transitions. Figure 2 compares the full regulariser with the baseline, Figure 3 varies hyperparameters, and Appendix C shows final samples and nearest neighbours. These results support practical gains, but do not directly test whether the proposed alignment reduces errors along the actual reverse trajectory, as claimed in Section 4.4. Evidence connecting the regulariser to this specific effect is needed.
Questions for authors
-
Score construction.
Which density and coordinate space define the score represented by Eq. 9 with the encoder-based kernel? How does its raw-coordinate weighted displacement follow from the latent-space score or its gradient through the encoder?
-
Connection to the stated sampling problem.
What specific approximation or estimation error causes the distortion described in Section 3.1, and what evidence shows that the combined objective reduces it? How does the forward-corruption bound in Appendix A.3 support this correction during actual reverse sampling?