acceptodds
Under review as a conference paper at ICLR 2027

Fine-Grained Graph Generation with Latent Mixture Scheduling

Abstract

Controlled graph generation requires producing graphs whose structures satisfy given topological attributes, instead of merely matching an unconditional graph distribution. Yet conditional latent-variable models can suffer from a mismatch between graph-conditioned latent representations used during training and attribute-conditioned representations used at inference. We introduce TOPOGEN, a conditional stochastic graph generator for fine-grained structural control. Its central component is a latent mixture scheduler that gradually transitions training from graph-conditioned training to attribute-only generation by interpolating latent distributions. This schedule reduces the mismatch between posterior-based reconstruction during training and prior-based conditional generation at inference time, enabling TOPOGEN to generate graphs from attributes alone. The approach allows improving joint constraint satisfaction without sacrificing diversity among valid outputs. Across five graph benchmark collections, TOPOGEN achieves the lowest average paired attribute error on all datasets and the best average structural fidelity across spectral and edit-distance metrics. Ablations show that latent scheduling and distribution alignment are important for controllable generation.

Primary area:
generative models
Keywords:
Controllable Graph Generation

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_zxEjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a graph generation method which allows conditioning on graph level topological attributes, such as number of maximum cliques, diameter and density. Many of such attributes are discrete, making them non-differentiable w.r.t. adding edges. This makes it difficult to enforce them directly through a loss. This paper takes a different approach by learning a graph-conditioned and attribute-conditioned latent distribution, which are progressively aligned using Wasserstein latent alignment. This approach makes it possible to include a wide set of conditioning attributes, while the downside of the latent alignment approach is that the attributes are soft conditioning signals, so there is no direct supervision signal from attributes during graph generation. The graph is constructed with a one-shot adjacency-based decoder. Expeirments show that the method produces graphs with similar attribute vectors, while structural evaluation metrics are mixed.

Critical strengths

The topic of conditional graph generation is very relevant and there are no great solutions for this problem yet. Most graph generation approaches do not allow conditioning on graph properties. Latent alignment is an interesting approach to this problem and it does not require the attributes to be differentiable.

Critical weaknesses

W1: Equation (11) does not specify how the graph reconstruction loss is computed. Based on (8) I assume it is an edgewise cross-entropy loss between the predicted and ground truth adjacency matrices. This means the loss assumes a fixed node ordering. However, the graph encoder uses a GNN and mean pooling making it permutation invariant. Thus, a permutation of a ground truth graph gives the same graph latent but differs in the adjacency matrix target. Graph-level graph generation typically addresses this with some kind of graph matching or fixed node ordering (not available for all graph families). It is unclear how this problem is resolved here.

W2: The mechanism for controlling graph size is indirect. The decoder works with a fixed-size adjacency matrix, and post-processing keeps the largest connected component. Thus, the requested size is only achieved if the decoder happens to produce a largest connected component with the given number of nodes. More targeted experiments would be necessary to establish that this mechanism actually reliably controls graph size. Also, it is unclear how unused nodes are handled in the reconstruction loss.

W3: Structural fidelity results are mixed, with improvements over the baselines only in some settings. This is attributed to a tradeoff between structural fidelity and attribute controllability. However, it could be that the simple adjacency-based decoder is fundamentally limited and cannot produce more complex graph structures.

W4: The number of seeds is not reported and standard deviations are not shown.

Questions for authors

Did you try interpolating or extrapolating in the attribute space, rather than using only attribute vectors computed from existing graphs? This would be the strongest application in my view.

To interpret the attribute-wise performance in Table 11, one (or both) of the following could be helpful to get a sense of scale for the reported numbers

  • report average and standard deviation of each given attribute for graphs in a given dataset
  • report the attribute-wise performance for some basline model(s)
8 comments
AuthorsOpenReview ↗

Response 1

Q: Did you try interpolating or extrapolating in the attribute space, rather than using only attribute vectors computed from existing graphs? This would be the strongest application in my view.

A: Our existing OOD experiment (Table 3) evaluates control vectors outside the training distribution, and these controls are extracted from generated Barabási–Albert graphs. We also evaluated TopoGen using synthetic control vectors formed by interpolating and extrapolating pairs of held-out control vectors. For each dataset, we randomly selected 25 distinct pairs of test control vectors ((ca,cb))((c_a,c_b)). For each pair, we constructed (c(α)=(1−α)ca+αcb)(c(\alpha)=(1-\alpha)c_a+\alpha c_b), with $(\alpha={0.25,0.5,0.75}) $ for interpolation and (α=−0.25,1.25)(\alpha={-0.25,1.25}) for extrapolation, giving 75 interpolation and 50 extrapolation candidates per dataset. Before generation, we rejected structurally infeasible controls, including invalid node/edge counts, bounded attributes outside ([0,1]), and inconsistent node-edge-density combinations. Across the five datasets, 359/375 (95.7%) interpolation controls and 168/250 (67.2%) extrapolation controls passed these checks. The macro-average MAD is 4.58 for interpolation and 4.69 for extrapolation across the five datasets. These results show that TOPOGEN can generate graphs from control vectors at previously unseen points in the attribute space. Interpolated controls are substantially more likely to remain structurally feasible, while extrapolation is more challenging because moving outside the observed attribute space can produce invalid combinations of graph properties.

DatasetInterp. ValidInterp. MAD ↓Extrap. ValidExtrap. MAD ↓
Citeseer74/75 (98.7%)1.8227/50 (54.0%)2.58
Ogbn-Arxiv72/75 (96.0%)3.9022/50 (44.0%)5.12
WordNet63/75 (84.0%)4.1227/50 (54.0%)4.85
MUTAG75/75 (100%)8.8350/50 (100%)6.64
MOLBACE75/75 (100%)4.2142/50 (84.0%)4.25
Average95.7%4.5867.2%4.69

Q: The number of seeds is not reported and standard deviations are not shown.

A: Below are the results of TOPOGEN over 3 seeds with updated implementation to reduce memory with the rest of the model as is and have updated the result table in the paper.

DatasetMAD (Mean ± Std)SD (Mean ± Std)GED (Mean ± Std)
MOLBACE2.68 ± 0.150.14 ± 0.0125.41 ± 0.88
MUTAG1.62 ± 0.210.18 ± 0.0416.51 ± 1.70
Citeseer1.66 ± 0.070.27 ± 0.0118.78 ± 0.21
Ogbn-Arxiv1.87 ± 0.010.43 ± 0.0045.12 ± 0.32
WordNet1.87 ± 0.010.53 ± 0.0126.61 ± 0.23

Q: Structural fidelity results are mixed, with improvements over the baselines only in some settings. This is attributed to a tradeoff between structural fidelity and attribute controllability. However, it could be that the simple adjacency-based decoder is fundamentally limited and cannot produce more complex graph structures.

A: We agree that decoder capacity can affect structural fidelity. To address concerns about the quadratic edge decoder and evaluating on large graphs, we reran TopoGen with an implementation that reduces intermediate memory during pairwise edge prediction and extended evaluation to larger graphs. The decoder still scores all node pairs and therefore remains (O(Nmax⁡2))(O(N_{\max}^2)), but avoids explicitly materializing an (Nmax⁡×Nmax⁡×H)(N_{\max}\times N_{\max}\times H) tensor. With these reruns, TopoGen achieves the lowest SD on Citeseer, Ogbn-Arxiv, and MUTAG, and the second-lowest on MOLBACE, with an average SD of 0.304, comparable to the best baseline average of 0.30, while retaining the strongest attribute controllability. We also observe stable SD of 0.46 on Ogbn-Arxiv up to 200 nodes and SD of 0.02 on SBM graphs up to 1000 nodes. These results suggest that the earlier mixed fidelity is not solely due to a fundamental limitation of the adjacency-based decoder, although the (O(Nmax⁡2))(O(N_{\max}^2)) scaling remains a limitation.

AuthorsOpenReview ↗

Response 2

Q: The mechanism for controlling graph size is indirect. The decoder works with a fixed-size adjacency matrix, and post-processing keeps the largest connected component. Thus, the requested size is only achieved if the decoder happens to produce a largest connected component with the given number of nodes. More targeted experiments would be necessary to establish that this mechanism actually reliably controls graph size. Also, it is unclear how unused nodes are handled in the reconstruction loss.

A: The requested node count enters the model through the attribute vector. The decoder outputs a fixed (Nmax⁡×Nmax⁡)(N_{\max}\times N_{\max}) graph, and the reconstruction loss is computed over the full fixed-size adjacency matrix; padded/unused nodes therefore correspond to zero rows and columns and are supervised as non-edges. The true node count is not provided to the decoder or reconstruction loss. After retaining the largest connected component, we obtain:

DatasetMean Target NodesMean Generated NodesMean Abs. Node Error ↓
WordNet17.7513.923.91
Ogbn-Arxiv20.0219.022.05
MUTAG17.6715.224.00
MOLBACE32.2726.666.26
Citeseer19.3718.061.87
Average——3.62

Q: To interpret the attribute-wise performance in Table 11, one (or both) of the following could be helpful to get a sense of scale for the reported numbers report average and standard deviation of each given attribute for graphs in a given dataset report the attribute-wise performance for some baseline model(s)

A: Below are the tables for attribute-wise performance of topogen compared with baseline models. The results show consistent improvements in node count, edge count, and number of cliques, for which TopoGen achieves the lowest errors across all five datasets. The gains are not uniform across all controls. Some baselines obtain lower errors for clustering, connectivity, transitivity, treewidth estimates, or diameter.

MUTAG:

AttributeDiGressGraphRNNGenStatEDGETopoGen
Density0.660.050.040.060.02
Edges17.677.008.339.223.78
Nodes15.006.005.676.332.78
Node Connectivity0.000.000.000.000.00
Avg. Clustering0.000.000.000.250.06
Closeness Centrality0.580.080.070.080.06
Number of Local Bridges4.113.563.782.443.67
Transitivity0.000.000.000.280.08
Edge Connectivity0.000.000.000.000.00
Number of Cliques17.677.008.339.113.78
Treewidth Min Degree1.110.330.780.670.56
Diameter7.893.113.113.112.22
MAD / Avg. Error5.392.262.512.631.42
SD0.880.210.210.280.14

Molbace:

AttributeDiGressGraphRNNGenStatEDGETopoGen
Density0.550.020.020.030.02
Edges32.609.228.657.876.57
Nodes28.819.018.197.005.76
Node Connectivity0.000.000.000.000.00
Avg. Clustering0.010.010.010.180.07
Closeness Centrality0.570.040.040.060.05
Number of Local Bridges11.119.427.477.896.08
Transitivity0.000.010.010.170.08
Edge Connectivity0.000.000.000.000.00
Number of Cliques32.469.388.788.537.51
Treewidth Min Degree1.030.100.080.780.41
Diameter12.423.513.434.413.54
MAD / Avg. Error9.963.393.063.082.51
SD0.910.120.120.140.13

Citeseer:

AttributeDiGressGraphRNNGenStatEDGETopoGen
Density0.550.120.130.140.07
Edges27.4024.3525.0626.955.80
Nodes17.7714.3614.2413.152.04
Node Connectivity0.410.000.000.000.03
Avg. Clustering0.230.200.170.260.17
Closeness Centrality0.520.110.090.120.07
Number of Local Bridges7.957.555.964.314.55
Transitivity0.200.180.170.190.15
Edge Connectivity0.410.010.040.000.03
Number of Cliques18.1816.3116.1811.564.64
Treewidth Min Degree2.171.401.131.361.10
Diameter3.901.710.971.711.35
MAD / Avg. Error6.645.525.354.981.67
SD0.850.430.410.450.28
AuthorsOpenReview ↗

Response 3

Arxiv:

AttributeDiGressGraphRNNGenStatEDGETopoGen
Density0.690.120.120.130.06
Edges29.1724.4726.8829.525.25
Nodes18.0813.1213.7213.761.67
Node Connectivity0.150.010.020.020.09
Avg. Clustering0.270.220.200.240.21
Closeness Centrality0.490.110.100.110.08
Number of Local Bridges8.106.866.835.305.31
Transitivity0.160.150.140.150.13
Edge Connectivity0.150.020.020.030.10
Number of Cliques18.5114.8115.9613.915.57
Treewidth Min Degree1.951.501.591.681.35
Diameter2.331.440.881.461.77
MAD / Avg. Error6.675.245.545.531.80
SD0.740.470.480.480.43

Wordnet:

AttributeDiGressGraphRNNGenStatEDGETopoGen
Density0.660.120.100.130.09
Edges15.9911.8112.7412.093.01
Nodes15.2411.2011.9411.233.75
Node Connectivity0.020.000.000.000.02
Avg. Clustering0.040.050.070.080.21
Closeness Centrality0.370.080.080.080.07
Number of Local Bridges13.9210.4711.1910.378.99
Transitivity0.020.020.030.040.18
Edge Connectivity0.020.000.000.000.02
Number of Cliques15.0411.1211.8811.143.63
Treewidth Min Degree0.330.370.450.440.99
Diameter1.210.820.830.911.54
MAD / Avg. Error5.243.844.113.881.87
SD0.380.320.320.330.54
Reviewer_uBqVOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The manuscript introduces TopoGen, a new generative ML model for conditioned generation of graphs. TopoGen relies on a latent space encoding of the graph structure and the desired property vector.

Critical strengths
  1. Attribute reliability TopoGen produces graphs with the requested attributes reliably.
  2. Complex attributes TopoGen learns to generate graphs with complex global properties such as node connectivity.
  3. Clear methodology The methodology is clearly presented.
Critical weaknesses
  1. High spectral distance TopoGen tends to have higher spectral distance than most of the baselines.
  2. Graph edit distance I do not understand how the GED metric is computed and what its purpose is.
  3. Scalability concerns TopoGen is tested on relatively small graphs (2-3 k-hop ego subgraphs). Further, appendix A.2 the TopoGen performance seems to be inversely related with the size of the graph. This is a strong weakness which I believe is understated in the main content.
  4. Unified MAD the provided MAD (Table 1) is the normalized average over all the attributes. It is possible that TopoGen outperforms the baseline on some of the metrics and underperforms on others.

Minor

The paper contains some typos and writing issues. I will try to list some of them here.

  1. l.74 Gto -> G to
  2. Sec 3 inconsistencies between Z/Z_G/Z_c and bolded Z/Z_G/Z_c
  3. l. 221 missing parenthesis after p_psi(G | Z
Questions for authors
  1. Could the authors please give more insights on the results of Table 8? How does TopoGen’s MAD evolve when tested on larger graphs?
  2. In Table 11, could the authors please report the size of the generated graphs? This helps in better understanding the magnitude of the provided MAD.
  3. Is TopoGen capable of generating graphs with defined eigenvalues? If not, what do you think is the model missing in order to achieve that?
  4. In Table 1 TopGen is shown to achieve higher SD than the baselines. Do you think this could possibly be related with TopoGen better abiding by the required characteristics? If so, could you show this with numerical experiments?
  5. How do you compute the GED metric? What is the reference graph here?
  6. What is the runtime and memory cost of TopoGen?
  7. Could the authors please provide the full MAD table over all the attributes and baselines?

I encourage the authors to focus on question 1, which is my major point of concern.

Please consider question 3 as an opportunity to discuss the possible scientific impact of your work rather than a limitation.

Finally, I encourage the authors to polish the writing before an eventual camera-ready version.

9 comments
AuthorsOpenReview ↗

Response 1

Thank you for pointing out the typos, notation, and missing parenthesis. We have corrected them and polished the writing in the recently uploaded version.

Q: Could the authors please give more insights on the results of Table 8? How does TopoGen’s MAD evolve when tested on larger graphs? Scalability concerns TopoGen is tested on relatively small graphs (2-3 k-hop ego subgraphs). Further, appendix A.2 the TopoGen performance seems to be inversely related with the size of the graph. This is a strong weakness which I believe is understated in the main content.

A: To directly address the scalability concerns, we re-ran TopoGen while making two design changes: (i) replacing the graph-size-dependent spatial latent with a fixed 128-dimensional latent, and (ii) replacing explicit pairwise tensors of size B×Nmax⁡×Nmax⁡×dB \times N_{\max} \times N_{\max} \times d with matrix-form bilinear scoring, $HWH^T $, where $ H \in \mathbb{R}^{N_{\max} \times d} $ denotes the node representations and $ W \in \mathbb{R}^{d \times d} $ is a learned weight matrix. These changes substantially reduce parameter growth and intermediate memory while retaining all-pairs edge prediction. They enabled evaluation on substantially larger graphs and also improved performance on the original benchmarks.

DatasetMax Node SizeMADSD
Ogbn-Arxiv602.100.437
Ogbn-Arxiv803.330.447
Ogbn-Arxiv1003.940.443
Ogbn-Arxiv2004.680.466
SBM100022.170.024

On Arxiv, SD remains stable as graph size increases from 60 to 200 nodes (0.437–0.466), while TopoGen continues to maintain meaningful attribute control at larger scales. The SBM result further shows that the updated implementation can scale to graphs with up to 1000 nodes while preserving strong global spectral structure. These results show that these implementation refinements extend TopoGen to large graph sizes and provide a practical basis for further scaling.

AuthorsOpenReview ↗

Response 2

Below are the results per attribute.

Ogbn-Arxiv, max node = 60:

AttributeDifference AvgGround Truth AvgPred Avg
Density0.050.190.22
Edges6.3136.7436.84
Nodes1.7222.9922.01
Node Connectivity0.071.011.06
Avg Clustering0.230.290.15
Closeness Centrality0.080.490.44
Number of Local Bridges5.659.884.63
Transitivity0.140.170.16
Edge Connectivity0.081.011.07
Number of Cliques7.2723.3429.19
Treewidth Min Degree1.663.074.43
Diameter1.993.305.19
Average2.108.548.78

Ogbn-Arxiv, max node = 80:

AttributeDifference AvgGround Truth AvgPred Avg
Density0.050.190.20
Edges9.4964.2964.43
Nodes1.5029.3329.14
Node Connectivity0.131.011.13
Avg Clustering0.300.410.16
Closeness Centrality0.080.480.44
Number of Local Bridges5.298.123.31
Transitivity0.160.270.16
Edge Connectivity0.151.011.14
Number of Cliques17.5932.1048.98
Treewidth Min Degree3.534.547.77
Diameter1.733.415.01
Average3.3312.1013.49

Ogbn-Arxiv, max node = 100:

AttributeDifference AvgGround Truth AvgPred Avg
Density0.040.190.19
Edges10.8378.7876.43
Nodes1.7433.4933.63
Node Connectivity0.101.011.08
Avg Clustering0.320.420.14
Closeness Centrality0.070.480.43
Number of Local Bridges5.378.543.69
Transitivity0.180.280.15
Edge Connectivity0.101.021.09
Number of Cliques22.6237.5059.47
Treewidth Min Degree4.244.998.87
Diameter1.773.455.14
Average3.9514.1815.86

Ogbn-Arxiv, max node = 200:

AttributeDifference AvgGround Truth AvgPred Avg
Density0.040.150.16
Edges15.2095.5992.36
Nodes3.5347.4047.07
Node Connectivity0.031.011.02
Avg Clustering0.240.300.10
Closeness Centrality0.090.460.39
Number of Local Bridges9.4417.308.42
Transitivity0.110.150.11
Edge Connectivity0.031.011.03
Number of Cliques20.4160.1979.11
Treewidth Min Degree4.574.749.07
Diameter2.603.455.99
Average4.6919.3120.40

SBM, max node = 1000:

AttributeDifference AvgGround Truth AvgPred Avg
Density0.000.010.01
Edges88.00657.70691.90
Nodes34.70436.70471.40
Node Connectivity0.001.001.00
Avg Clustering0.010.010.01
Closeness Centrality0.020.170.18
Number of Local Bridges40.20101.30128.30
Transitivity0.000.010.01
Edge Connectivity0.001.001.00
Number of Cliques86.30648.40683.70
Treewidth Min Degree15.2052.0053.40
Diameter1.7014.0013.90
Average22.18159.36170.40
AuthorsOpenReview ↗

Response 3

Q: In Table 11, could the authors please report the size of the generated graphs? This helps in better understanding the magnitude of the provided MAD.

A: All datasets in Table 11 use (Nmax⁡=50)(N_{\max}=50), while the generated graph size can be at most 50 determined by conditioning on control attributes. The following Table provides additional context for the reported node-count errors, reporting the mean target size, mean generated size, and mean absolute node-count error:

DatasetMean Target NodesMean Generated NodesMean Abs. Node Error
WordNet17.7513.923.91
Ogbn-Arxiv20.0219.022.05
MUTAG17.6715.224.00
MOLBACE32.2726.666.26
Citeseer19.3718.061.87

Q: Is TopoGen capable of generating graphs with defined eigenvalues? If not, what do you think is the model missing in order to achieve that?

A: We thank the reviewer for their thoughtful question. We believe the scientific opportunity here is to study joint control of global spectral structure and explicit topological properties, which is quite interesting. The research challenge is to determine that the spectral targets and structural attributes are jointly feasible. For example, for the combinatorial Laplacian, the sum of the eigenvalues equals twice the number of edges. Consequently, a prescribed spectrum and an independently requested edge count cannot always be satisfied together. Spectre (Martinkus et al. 2022) generates graphs using Laplacian eigenvalues and eigenvectors.

In terms of empirical experiments, we additionally tested whether TopoGen can directly condition on spectral properties by replacing its attribute controls with the first five non-trivial normalized-Laplacian eigenvalues, (λ2,…,λ6)(\lambda_2,\ldots,\lambda_6), using the TopoGen architecture as is. We experimented on Citeseer dataset and TopoGen achieves an average spectral MAD of 0.229:

ControlTarget Avg.Generated Avg.MAD ↓
λ2\lambda_20.1730.1750.106
λ3\lambda_30.3920.3630.196
λ4\lambda_40.5870.5280.257
λ5\lambda_50.7510.6870.298
λ6\lambda_60.9130.8280.290
Average0.5630.5160.229

These results show that TopoGen can approximately control selected spectral characteristics, not only the hand-designed structural attributes used in the main experiments. We do not claim exact realization of an arbitrary full spectrum; more precise spectral matching could potentially be obtained by directly supervising the generated eigenvalues.

Reviewer_6YWGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TopoGen, a conditional graph generation framework designed to generate graphs satisfying fine-grained topological constraints. The key contribution is a latent mixture scheduler that gradually transitions training from graph-conditioned posterior representations to attribute-conditioned prior representations, reducing the training–inference mismatch. The framework also incorporates Wasserstein distribution alignment and attribute reconstruction objectives. Experiments across five graph benchmarks demonstrate improvements in attribute controllability, supported by structural fidelity evaluations and ablation studies.

Critical strengths

(1) The paper addresses the important problem of fine-grained controllable graph generation, particularly the mismatch between graph-conditioned training and attribute-only inference.

(2) The proposed latent mixture scheduler provides a simple and intuitive mechanism for gradually aligning training and inference distributions, supported by a well-defined Gaussian interpolation formulation.

(3) The experiments cover diverse graph datasets and evaluate multiple aspects of generation quality, including attribute satisfaction, structural fidelity, novelty, diversity, and out-of-distribution generalization.

Critical weaknesses

(1) The methodological novelty is somewhat limited, as the framework primarily combines Gaussian latent interpolation, Wasserstein alignment, and attribute reconstruction, without introducing a fundamentally new graph generation architecture.

(2) Although TopoGen achieves the best average attribute error, its structural fidelity results are less convincing. In Table 1, its average spectral difference is worse than both GenStat and GraphRNN, contradicting the abstract's claim of the best average structural fidelity.

(3) The evaluation is limited to undirected, unlabeled graphs and extracted local subgraphs. Furthermore, the quadratic edge decoder raises scalability concerns, limiting the demonstrated applicability to large-scale and realistic graph generation tasks.

Questions for authors

Please discuss the trade-off between attribute controllability and structural fidelity.

4 comments
AuthorsOpenReview ↗

Response 1

Q: The evaluation is limited to undirected, unlabeled graphs and extracted local subgraphs.Furthermore, the quadratic edge decoder raises scalability concerns, limiting the demonstrated applicability to large-scale and realistic graph generation tasks.

A: We focus on a novel training—inference mechanism that, in contrast to many existing approaches for controlled graph generation, is scalable. To illustrate, we extend the evaluation to graphs with up to 1000 nodes. Extending TopoGen to directed and labeled graphs is a natural next step; the latent conditioning and scheduling framework is not inherently restricted to unlabeled graphs, while such settings would require corresponding extensions to the graph decoder.

To illustrate scalability, we reran TopoGen with a significantly larger number of nodes. To achieve this scalability level, we considered two design choices in our implementation: replacing the commonly-used graph-size-dependent spatial latent with a fixed 128-dimensional latent, and replacing explicit pairwise tensors of size B×Nmax⁡×Nmax⁡×dB \times N_{\max} \times N_{\max} \times d with matrix-form bilinear scoring, $HWH^T $, where $ H \in \mathbb{R}^{N_{\max} \times d} $ denotes the node representations and $ W \in \mathbb{R}^{d \times d} $ is a learned weight matrix. These changes substantially reduce parameter growth and intermediate memory while retaining all-pairs edge prediction. This enabled evaluation on substantially larger graphs and also improved performance on the original benchmarks. Updated TopoGen results, including the large-graph evaluations, are reported below:

MetricModelWordNetCiteseerOgbn-ArxivMUTAGMOLBACEAvg.
MAD ↓Erdős-Rényi5.996.957.166.1811.057.46
GraphRNN3.26‾\underline{3.26}5.054.80‾\underline{4.80}1.71‾\underline{1.71}3.813.73‾\underline{3.73}
GenStat4.115.345.534.143.05‾\underline{3.05}4.43
EDGE3.914.97‾\underline{4.97}5.522.623.074.02
DiGress5.236.636.675.399.966.78
GruM3.755.375.423.8010.805.83
TopoGen1.871.641.871.412.501.86
SD ↓Erdős-Rényi0.980.891.010.660.490.80
GraphRNN0.310.420.46‾\underline{0.46}0.21‾\underline{0.21}0.130.31
GenStat0.32‾\underline{0.32}0.40‾\underline{0.40}0.470.21‾\underline{0.21}0.110.30
EDGE0.32‾\underline{0.32}0.440.470.270.140.33
DiGress0.370.850.730.880.910.75
GruM0.400.430.500.520.690.51
TopoGen0.540.280.430.140.13‾\underline{0.13}0.30
GED ↓Erdős-Rényi34.2446.3749.1336.0066.3246.41
GraphRNN32.5854.8352.4315.77‾\underline{15.77}41.52‾\underline{41.52}39.43
GenStat37.0158.6963.7734.6661.3251.09
EDGE35.1659.5859.7029.7772.9351.43
DiGress31.3245.16‾\underline{45.16}47.2432.6661.6443.60
GruM29.6855.0651.8727.0065.2745.78
TopoGen26.3540.4644.9814.5524.4430.15
DatasetMax Node SizeMADSD
Ogbn-Arxiv602.100.437
Ogbn-Arxiv803.330.447
Ogbn-Arxiv1003.940.443
Ogbn-Arxiv2004.680.466
SBM100022.170.024

On Arxiv, SD remains stable as graph size increases from 60 to 200 nodes (0.437–0.466), while TopoGen continues to maintain meaningful attribute control at larger scales. The SBM result further shows that the updated implementation can scale to graphs with up to 1000 nodes while preserving strong global spectral structure. These results show that these implementation refinements extend TopoGen to large graph sizes and provide a practical basis for further scaling.

AuthorsOpenReview ↗

Response 2

Q: The methodological novelty is somewhat limited, as the framework primarily combines Gaussian latent interpolation, Wasserstein alignment, and attribute reconstruction, without introducing a fundamentally new graph generation architecture.

A: Our methodological contribution is the training mechanism that addresses the mismatch between graph-conditioned reconstruction during training and attribute-only generation at inference. Specifically, we progressively shift the decoder’s training input from graph-posterior latents to attribute-prior latents, ending in prior-only training that matches the inference setting. A practical advantage is that this procedure does not require differentiating through graph-attribute computations or using an external deterministic tool to compute graph attributes during training. The ablations support the contribution of this training mechanism: removing either the scheduler or Wasserstein alignment increases MAD on all three datasets reported in Table 6. We will highlight this contribution statement in the paper to emphasize the scheduled transition to attribute-only generation and the flexibility of working with non-differentiable graph attributes.

Q: Although TopoGen achieves the best average attribute error, its structural fidelity results are less convincing. In Table 1, its average spectral difference is worse than both GenStat and GraphRNN, contradicting the abstract's claim of the best average structural fidelity.

A: With the updated implementation, TopoGen achieves substantially improved structure fidelity: it obtains the best SD on Citeseer (0.28), Ogbn-Arxiv (0.43), and MUTAG (0.14), second-best on MOLBACE (0.12), and an average SD of 0.30, comparable to the best baseline (0.30). At the same time, its average MAD improves from 3.32 to 1.86, achieving the best controllability across all five datasets. We have revised the abstract to avoid claiming best structural fidelity and updated the paper with revised results.

Q: Please discuss the trade-off between attribute controllability and structural fidelity.

A: Controllability and reference based structural fidelity measure complementary aspects of generation. MAD measures whether the generated graph satisfies the requested control attributes, whereas SD measures similarity to a particular held-out reference graph. A single control vector does not uniquely specify a single graph, in fact, multiple non-isomorphic graphs can satisfy the same attributes while having different spectra. Thus, strong controllability does not necessarily imply minimum distance to one reference graph. TopoGen achieves the lowest overall MAD while maintaining competitive SD. In fact, on WordNet it substantially reduces error on several attributes despite a higher SD. These results suggest that stronger adherence to the requested graph characteristics can generally be achieved without sacrificing spectral fidelity, while also highlighting that the two metrics capture complementary aspects of graph generation.

Thank you again for your review. We wanted to follow up on our November 13 response. If any of your concerns are still open, we would be grateful for the opportunity to clarify them or address any questions during the discussion period.

Thank you for the detailed rebuttal and additional experiments. However, I remain concerned about the limited methodological novelty and the applicability of the approach to more realistic settings.

Reviewer_boT3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces TOPOGEN, a conditional stochastic graph generator for fine-grained structural control. Its central component is a latent mixture scheduler that gradually transitions training from graph-conditioned training to attribute-only generation by interpolating latent distributions. This schedule reduces the mismatch between posterior-based reconstruction during training and prior-based conditional generation at inference time, enabling TOPOGEN to generate graphs from attributes alone.

Critical strengths

S1.This paper points out the discrepancy between the "graph-based representation" during the training phase and the "attribute-based representation" used during the inference phase. It motivates the paper very well.

S2. TOPOGEN achieves the lowest MAD on four out of the five datasets, as well as the lowest average attribute error across all datasets, providing strong experimental evidence.

Critical weaknesses

W1. the attribute reconstruction loss constrains the retention of control information within the latent representation, rather than the actual attributes of the final generated graph. There is no direct guarantees for the joint satisfaction of attribute constraints in generated graphs. W2. The paper uses the scheduling coefficient \beta to control the transition from the graph-conditioned posterior to the attribute-conditioned prior. This process seems fixed and heuristic.

Questions for authors
  1. Does TOPOGEN offer any theoretical guarantees regarding the simultaneous satisfaction of multiple target attributes? If not, could you clarify the scope to which your conclusions about controllability apply? Please report the proportion of generated graphs that satisfy all attribute constraints simultaneously
  2. Could the authors clarify how the transition from the graph-conditioned posterior to the attribute-conditioned in practice?
4 comments
AuthorsOpenReview ↗

Response 1

Q: The attribute reconstruction loss constrains the retention of control information within the latent representation, rather than the actual attributes of the final generated graph. There is no direct guarantees for the joint satisfaction of attribute constraints in generated graphs. Does TOPOGEN offer any theoretical guarantees regarding the simultaneous satisfaction of multiple target attributes? If not, could you clarify the scope to which your conclusions about controllability apply? Please report the proportion of generated graphs that satisfy all attribute constraints simultaneously

A: We agree with the reviewer that TOPOGEN does not provide a theoretical guarantee of simultaneous satisfaction of all target attributes. The attribute reconstruction loss encourages the latent representation to retain the requested control information. Our controllability claims are therefore empirical: we measure how closely the estimated structural attributes of generated graphs match the requested targets. We will clarify this point in the paper. We would like to highlight that our contribution is in the training mechanism that does not require differentiating through graph-attribute computations or using an external deterministic tool to compute graph attributes during training.

To directly evaluate simultaneous constraint satisfaction, we evaluate whether all target attributes fall within a specified tolerance of their respective targets. Let (cij)(c_{ij}) denote the requested value of attribute (j) for evaluation example (i), and let (c^ij)(\hat c_{ij}) denote the corresponding value computed from the generated graph (G^i)(\hat G_i). To account for differences in attribute scales, we obtain the minimum and maximum raw values of each attribute from the training set and use the same statistics for both target and generated attributes. For an attribute with positive training range (rj=cjmax⁡−cjmin⁡)(r_j=c_j^{\max}-c_j^{\min}), the normalized error is ∣c^ij−cij∣rj.\frac{|\hat c_{ij}-c_{ij}|}{r_j}. A generated graph satisfies the joint criterion only if every evaluated attribute has normalized error at most (τ)(\tau). We report the percentage of generated graphs meeting this criterion: $\frac{100}{N} \sum_{i=1}^{N} \mathbf{1} \left[ \max_{j=1,\ldots,12} e_{ij}\leq\tau \right],$ where (N) is the number of evaluated target–generated-graph pairs. We report results for (τ∈0.05,0.10,0.20,0.30,0.40,0.50\tau\in{0.05,0.10,0.20,0.30,0.40,0.50}), i.e., errors within 5–50% of each attribute's training-set range. The resulting JointSat rates are

Dataset5%10%20%30%40%50%
Citeseer0.002.568.9714.1038.4642.31
MUTAG0.000.000.000.0011.1133.33
Molbace0.000.006.766.768.1121.62
Arxiv0.080.533.606.029.0136.24
WordNet0.170.651.612.977.8935.24

The results show that simultaneous satisfaction of all attributes is challenging. At the more permissive 50% tolerance, it ranges only from 21.62–42.31%. However, we note the sensitivity of these results: Consider the following 3-node graph: A — B — C Misgenerating only one edge in this graph significantly affects attribute values. For example, if we just add A — C to the above graph: average clustering and transitivity both increase from 0 to 1, diameter decreases from 2 to 1, and node and edge connectivity both increase from 1 to 2. These changes can exceed all tolerance values under normalization by any non-zero training-set range.

AuthorsOpenReview ↗

Response 2

Per-attribute satisfaction: this table reports the percentage of generated graphs whose realized value for each attribute (A) lies within the specified tolerance of the requested target:

DatasetTol.A0A1A2A3A4A5A6A7A8A9A10A11
Citeseer5%41.0388.4675.6498.7215.3819.2319.2317.9596.1571.7928.2123.08
Citeseer10%71.7996.1592.3198.7220.5135.9050.0030.7796.1589.7467.9523.08
Citeseer20%88.4697.44100.0098.7244.8770.5184.6250.0096.1597.4497.4423.08
Citeseer30%94.8798.72100.0098.7266.6787.1891.0369.2396.15100.0098.7223.08
Citeseer40%96.1598.72100.0098.7287.1893.5997.4480.7796.15100.00100.0055.13
Citeseer50%97.4498.72100.0098.7294.8797.4498.7288.4696.15100.00100.0055.13
MUTAG5%0.0011.110.00100.0044.440.0011.1144.44100.000.0055.5611.11
MUTAG10%11.1111.1111.11100.0044.4411.1133.3344.44100.0011.1155.5611.11
MUTAG20%33.3355.5655.56100.0044.4422.2255.5644.44100.0044.4455.5644.44
MUTAG30%55.5677.7877.78100.0044.4455.5666.6744.44100.0055.5655.5655.56
MUTAG40%88.8988.8988.89100.0044.4477.7888.8944.44100.0088.8955.5655.56
MUTAG50%100.00100.00100.00100.0044.4488.89100.0044.44100.0088.89100.0066.67
Molbace5%27.0332.4336.49100.0012.168.1136.4912.16100.0027.0343.242.70
Molbace10%45.9543.2456.76100.0013.5120.2748.6513.51100.0040.5443.2413.51
Molbace20%78.3872.9775.68100.0020.2747.3072.9713.51100.0059.4643.2447.30
Molbace30%85.1481.0882.43100.0024.3268.9291.8917.57100.0078.3843.2466.22
Molbace40%87.8490.5489.19100.0029.7382.4395.9522.97100.0082.4343.2486.49
Molbace50%91.8994.5994.59100.0036.4989.1998.6525.68100.0091.8993.2494.59
Arxiv5%55.2192.3970.7395.8315.9417.9536.9228.4495.3080.8031.9611.81
Arxiv10%77.5599.0990.1695.8324.6934.6559.7148.3995.3093.3072.4311.81
Arxiv20%90.50100.0098.3095.8347.2561.6886.9075.5095.3099.8192.7311.81
Arxiv30%95.53100.0099.8595.8366.9879.7094.3688.9495.30100.0095.6811.81
Arxiv40%97.08100.0099.9299.7783.9590.9598.1894.6299.81100.0098.7511.81
Arxiv50%98.33100.00100.0099.7793.6897.4399.5597.8099.81100.0099.7739.83
WordNet5%12.0068.5945.6998.9422.0115.3515.5820.5798.4357.5924.8817.46
WordNet10%39.8291.8066.4498.9429.8031.6833.9730.5598.4386.6024.8817.46
WordNet20%73.6899.2588.4598.9448.3964.5663.9455.4798.4399.4224.8817.46
WordNet30%87.9499.7998.1998.9465.3885.3080.2872.5698.4399.9324.8817.46
WordNet40%93.9299.9399.6998.9479.9495.1191.0885.2798.43100.0082.4017.46
WordNet50%96.27100.0099.9798.9488.9697.9596.9291.7698.43100.0082.4053.86

The per-attribute results show that control accuracy varies substantially across attributes. For example, at the 20% tolerance on Citeseer, six attributes individually satisfy their thresholds in more than 95% of generated graphs, while only 8.97% of graphs satisfy all 12 thresholds simultaneously. This shows the sensitivity of all attribute analysis.

AuthorsOpenReview ↗

Response 3

Q: The paper uses the scheduling coefficient \beta to control the transition from the graph-conditioned posterior to the attribute-conditioned prior. This process seems fixed and heuristic. Could the authors clarify how the transition from the graph-conditioned posterior to the attribute-conditioned in practice?

A: The scheduling coefficient β\beta controls the transition from the graph-conditioned posterior to the attribute-conditioned prior, with increasing β\beta progressively shifts the latent representation from the graph-conditioned posterior toward the attribute-conditioned prior. In the final phase, it is set to β\beta = 1, so the decoder is trained entirely from the same attribute-conditioned latent representation available at inference. This coefficient is not learned or adjusted online, but the interpolation has a precise geometric interpretation: increasing β\beta moves the decoder’s latent input distribution toward the inference prior along the Gaussian Wasserstein geodesic (Line 155—157). Our ablations also evaluate gradual versus constant prior inclusion and sensitivity to the scheduler parameters. In Table 6, removing the scheduler increases MAD from 1.64 to 1.87 on Citeseer, and 2.50 to 3.47 on MOLBACE. These results support the utility of this scheduling strategy, without claiming that its particular functional form is optimal.

AblationMUTAGMOLBACECiteSeer
TOPOGEN1.412.501.64
w/o Wasserstein alignment2.563.894.85
w/o Attribute Reconstruction1.583.411.68
w/o Mixture Scheduler1.113.471.87
w/o G, (Prior only)1.262.961.73
w/o c, (Posterior only)1.913.321.87

Thank you again for your review. We wanted to follow up on our November 13 response. If any of your concerns are still open, we would be grateful for the opportunity to clarify them or address any questions during the discussion period.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to all reviewers

We thank all the reviewers for their time and comments. Reviewers recognized the importance of fine-grained controllable graph generation, the training-inference mismatch addressed by TopoGen, and the empirical controllability demonstrated through evaluation of attribute satisfaction, structural fidelity, novelty and diversity of generated graphs, and out-of-distribution generalization. The reviewers also raised comments about edge decoder and the limited graph sizes considered in the experiments.

In response to the reviews, we ran TopoGen experiments for substantially larger graphs and reduced its memory usage. We also conducted additional analyses directly addressing individual reviewer comments about structural fidelity, simultaneous attribute satisfaction, and interpolation/extrapolation.

We have updated the paper with these results and clarified details of the implementation. We have uploaded a revised PDF in response to all four reviews; changed text is shown in blue.

Regarding the intermediate memory required for pairwise edge prediction: the decoder evaluates every node pair, and therefore remains (O(Nmax⁡2))(O(N_{\max}^2)). However, our updated implementation does not explicitly create a 3D tensor, Nmax⁡×Nmax⁡×HN_{\max}\times N_{\max}\times H(H: representation size) for node pairs. Thus, edge prediction is quadratic in Nmax⁡N_{\max} but the intermediate memory usage is substantially small. Our updated main results (Table 1 in the paper) with this change are as follows:

MetricModelWordNetCiteseerOgbn-ArxivMUTAGMOLBACEAvg.
MAD ↓Erdős-Rényi5.996.957.166.1811.057.46
GraphRNN3.26‾\underline{3.26}5.054.80‾\underline{4.80}1.71‾\underline{1.71}3.813.73‾\underline{3.73}
GenStat4.115.345.534.143.05‾\underline{3.05}4.43
EDGE3.914.97‾\underline{4.97}5.522.623.074.02
DiGress5.236.636.675.399.966.78
GruM3.755.375.423.8010.805.83
TopoGen1.871.641.871.412.501.86
SD ↓Erdős-Rényi0.980.891.010.660.490.80
GraphRNN0.310.420.46‾\underline{0.46}0.21‾\underline{0.21}0.130.31
GenStat0.32‾\underline{0.32}0.40‾\underline{0.40}0.470.21‾\underline{0.21}0.110.30
EDGE0.32‾\underline{0.32}0.440.470.270.140.33
DiGress0.370.850.730.880.910.75
GruM0.400.430.500.520.690.51
TopoGen0.540.280.430.140.13‾\underline{0.13}0.30
GED ↓Erdős-Rényi34.2446.3749.1336.0066.3246.41
GraphRNN32.5854.8352.4315.77‾\underline{15.77}41.52‾\underline{41.52}39.43
GenStat37.0158.6963.7734.6661.3251.09
EDGE35.1659.5859.7029.7772.9351.43
DiGress31.3245.16‾\underline{45.16}47.2432.6661.6443.60
GruM29.6855.0651.8727.0065.2745.78
TopoGen26.3540.4644.9814.5524.4430.15

These updated results show substantially improved attribute agreement and structural fidelity. We revised the paper to reflect these findings and document the implementation changes and experimental settings used for the updated evaluation.

We also incorporated the larger-graph experiments and the additional analyses of simultaneous attribute satisfaction and interpolation/extrapolation.

DatasetMax Node SizeMADSD
Ogbn-Arxiv602.100.437
Ogbn-Arxiv803.330.447
Ogbn-Arxiv1003.940.443
Ogbn-Arxiv2004.680.466
SBM100022.170.024

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.