- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes P SGP, which treats supervised graph prediction as conditional generation over attributed graph isomorphism classes. A permutation-equivariant flow model learns to reconstruct a clean graph from a corrupted graph while conditioning on an image, molecular fingerprint, or genotype sequence. Training uses coordinate-wise losses without aligning predicted and target graphs. At inference, the method generates one graph or selects the most frequent graph class among multiple samples using Weisfeiler–Lehman hashes.
Critical strengths
- The method combines permutation-equivariant conditional flow matching with coordinate-wise supervision, and Proposition 2 establishes that the expected training objective is invariant to arbitrary target reindexing. This provides a sound basis for predicting graphs with discrete or continuous attributes while avoiding the explicit alignment step required by matching-based approaches.
- A single generated candidate outperforms the strongest reported baseline on all six established benchmarks, with substantial gains on molecular reconstruction and USCities. Modal decoding improves performance further, and the appendix examines size prediction, clustering accuracy, and sampling budgets. These results support the practical usefulness of the approach across multiple input modalities and graph types.
Critical weaknesses
- The method closely builds on DiGress, DeFoG, and DIRECTO, but comparisons focus on matching-based predictors. Without a conditional diffusion baseline or a simpler DeFoG adaptation using comparable architectures and training budgets, it is difficult to determine which proposed design choices account for the gains. This limits the evidence for a distinct methodological advance despite the strong benchmark results.
- The genotype experiment compares only against a deterministic predictor, which necessarily incurs positive discrepancy when the reference distribution is multimodal. Furthermore, the constructed reference distributions are not the simulator’s full conditional posterior, and generic and designated non-identifiable examples can include the same family structures while receiving different reference distributions. These issues limit the interpretation of the reported MMD improvement as evidence of recovering conditional graph uncertainty; stronger probabilistic baselines and clarification of the statistical target are needed.
Questions for authors
- Do you have comparisons against a conditional graph diffusion model or a simpler DeFoG adaptation using comparable input encoders, graph backbones, and training budgets? These would help determine whether the gains depend on the proposed design choices.
- Appendix B.4 appears to allow the generic and non-identifiable branches to generate the same two-person, depth-four family structures, while assigning point-mass and uniform reference distributions, respectively. Is this interpretation correct? If so, how should the reference distribution be understood as a prediction target given only the observed genotype?
- How does P SGP compare with a probabilistic classifier over the benchmark’s finite family-template catalogue? Also, are the genotype MMD results stable when varying integration steps and replacing the terminal argmax with categorical sampling?
4 comments
Official Comment by Authors (2/2)
W2/Q2: Generic and non-identifiable branches. Yes, this interpretation is correct, and for these structures both reference distributions favor the same prediction. In the test set, 51 of the 10,000 examples are generic examples with the same structure as a non-identifiable two-person example. In both branches, each example consists of genotypes and the one family graph used to simulate them, and the model is trained only on these pairs. Given only the genotypes, the prediction target is therefore the simulator's posterior, that is, the conditional distribution of the family graph given the genotypes, which the model can only learn from how often each family graph occurs with similar genotypes during training. As Sec. 3.3 states, this posterior is intractable to evaluate exactly, so neither reference distribution equals it. For a generic example, the benchmark compares the prediction with the family graph used to simulate its genotypes, the only graph that the example contains, so the reference distribution is a point mass on that graph. The generic branch is not restricted to structures that the genotypes can distinguish, which is why we call it generic rather than identifiable. A generic example can therefore leave the family graph uncertain, but without the posterior we do not know which other structures its genotypes leave possible or how likely each one is, and a single graph per example cannot show how a model spreads its probability. We introduced the non-identifiable branch to test this uncertainty in a controlled way. It chooses with equal probability among structures that the genotypes are known not to distinguish, so these structures are equally likely given the genotypes, and its reference distribution spreads probability equally over them.
A two-person family in which the two individuals descend from one shared ancestor through four edges in total has three such structures, in which the shared ancestor is 1, 2, or 3 generations above the first individual and 3, 2, or 1 generations above the second, so the uniform reference distribution gives each of them probability 1/3. When the generic branch produces one of these three structures, it also chooses each with equal probability, so the point mass of such a generic example falls on one of the three structures, chosen uniformly at random, and averaging this point mass over the random choice gives exactly the uniform reference distribution. Both reference distributions therefore favor the same prediction, probability 1/3 on each structure, which has the lowest MMD² against the uniform reference distribution and the lowest average MMD² against the point-mass one. The overlap thus does not change which prediction the reported MMD² rewards for these structures. The revised App. B.4 includes this explanation.
Q3: Integration steps and terminal sampling. Yes, the conclusions are stable. In new evaluations (App. A.7), we drew two independent sets of 16 samples from the paper's checkpoint per validation input and averaged their MMD², either taking the most probable category at the final step (argmax) or sampling it (categorical):
| Steps | MMD², all, argmax | MMD², all, categorical | MMD², non-identifiable, argmax | MMD², non-identifiable, categorical |
|---|---|---|---|---|
| 100 | 0.2199 | 0.2231 | 0.3178 | 0.3491 |
| 250 | 0.2162 | 0.2134 | 0.3329 | 0.3100 |
| 500 | 0.2169 | 0.2180 | 0.3118 | 0.3175 |
| 1,000 | 0.2150 | 0.2170 | 0.3168 | 0.3025 |
At 250 steps, categorical minus argmax is -0.0027 overall (paired 95% CI -0.0105 to 0.0050) and -0.023 on the non-identifiable cases (-0.049 to 0.003), and both intervals include zero. On the non-identifiable cases every setting stays below 0.35, while any deterministic predictor scores at least 0.5 on each such case, so pSGP's advantage over deterministic prediction depends on neither the terminal step rule nor the number of steps.
Again, thank you for your valuable feedback. Given our response and clarifications to your questions and concerns, as well as the new results, we kindly ask you to consider increasing your score. We welcome further suggestions and are open to more questions.
I thank the authors for their efforts to address my concerns. My concerns were partially addressed. I would like to maintain my original scores.
Thank you for following up. Could you please clarify which of your concerns remain unresolved? We would like to address them if possible while the discussion period is still open.
Official Comment by Authors (1/2)
We thank the reviewer for the careful reading and recognizing the sound basis of Proposition 2, the gains on all six benchmarks from a single sample and further with modal decoding, and the usefulness across input modalities and graph types. We reply to the raised concerns next.
W1/Q1: Conditional diffusion and simpler DeFoG baselines. DiGress, DeFoG, and DIRECTO focus mainly on unconditional graph generation. Their conditional experiments use only graph-level labels or properties, such as a molecule's dipole moment in DiGress, the tertiary lymphoid structure content of a cell graph in DeFoG, or the execution time of a computational graph in DIRECTO. A simpler DeFoG adaptation therefore would condition only through the global graph features fed with a pooled embedding of the input. As new baselines, we trained this adaptation and a version of pSGP that replaces its flow with DiGress's diffusion process but keeps pSGP's full conditioning of the node and edge updates on the input tokens. Both use the same encoders, backbone size, frozen size predictors, training budgets, and argmax terminal step as pSGP. On validation inputs of Coloring 10 and of Genotypes:
The table separates the two design choices, the conditioning and the generator. The conditioning matters most on Coloring 10, where the simpler DeFoG adaptation is far below pSGP, and on GDB13, where our conditioning ablations (see App. A.5 and our response to 4Qhd) show that conditioning only through the global features lowers single GI-Acc from 68.8% to 22.0%. These results also show that pSGP works well with diffusion too. With DiGress's diffusion process in place of DeFoG's flow matching, it reaches 88.8% single and 97.9% modal@32 GI-Acc on Coloring 10, far above the simpler DeFoG adaptation (60.1% and 70.8%) and within 0.7 points of flow matching in modal@32 GI-Acc. On Genotypes, all three models perform similarly. The gains therefore come mainly from conditioning the node and edge updates on the input tokens, with either generator. These results support the point we made in the conclusion (Sec. 4): our proposed probabilistic formulation of SGP is not limited to DeFoG and can be instantiated with other suitable generative processes, such as discrete diffusion. Our main contribution is the formulation of matching-free probabilistic SGP, rather than a new flow-matching algorithm. We still chose to use DeFoG as our main instantiation because it disentangles sampling from training, and its paper reports better results with fewer sampling steps than most diffusion models. We added these results to App. A.6.
While running these comparisons, we also found and fixed a bug in the code used for the probabilistic Genotypes evaluation of pSGP. Training, the trained model, and the evaluations of the other benchmarks are not affected. With the fix, pSGP's Genotypes test MMD² is 0.2146 overall (0.225 in the submission) and 0.3284 on the non-identifiable cases (0.388 in the submission), and the revised Table 2 reports these values. All Genotypes results in this response use the fixed evaluation.
W2/Q3: Catalogue classifier. We trained the requested classifier as a new baseline, which encodes the genotype tokens with the same encoder architecture as pSGP and maps the result through a linear layer to a probability for each of the 20 graph isomorphism classes found among the training targets (the benchmark's catalogue). It receives neither the true graph size nor whether the example is generic or non-identifiable. On the test examples, it performs similarly to pSGP:
The overall difference, classifier minus pSGP, is -0.0018, and its paired 95% CI (-0.0047 to 0.0008) includes zero. It is important to note that the classifier works well here because all test graphs come from a small finite catalogue, but it cannot predict structures outside it or continuously varying attributes, whereas pSGP generates nodes and edges directly. Any deterministic predictor, such as Any2Graph, places all probability on one structure and therefore scores at least 0.5 on each non-identifiable case, and at least 0.558 on average over the 1,524 non-identifiable test cases, compared with 0.3284 for pSGP. pSGP thus performs on par with a classifier that has the catalogue built in, and together with the diffusion version of pSGP and the simpler DeFoG adaptation above, this classifier provides the stronger probabilistic baselines that the reviewer asked for. We added this to App. A.7.