acceptodds
Under review as a conference paper at ICLR 2027

G2SAE: Spectrally Supervised Graph-Level Autoencoding

Abstract

Graph-level autoencoding aims to compress graphs into graph-level latent representations from which their pairwise connectivity and node-level attributes can be recovered. Learning such representations requires preserving structural information while accounting for arbitrary node orderings during reconstruction supervision. We introduce G2SAE, a graph-level autoencoder that uses an explicitly supervised Laplacian spectral representation to guide graph reconstruction and node alignment. The encoder derives a spectral latent from graph topology and uses topological features to condition separate node and edge semantic latents. The decoder predicts node-level spectral coordinates from the spectral latent, then combines these coordinates and their pairwise relations with the semantic latents to reconstruct connectivity and attributes. The spectral representation also provides structural cues to align predicted and input nodes, enabling spectral and graph reconstruction objectives to supervise both the intermediate representation and the recovered graph. Experiments on molecular graph benchmarks demonstrate improved reconstruction fidelity, highlighting the effectiveness of G2SAE in learning graph-level representations that retain structural and semantic information.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph, AutoEncoder, Encoder-Decoder, Deep Learning, Representation Learning, Graph Matching

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_297qOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a graph-level autoencoder using explicit supervision of spectral coordinates. The authors claim that their method improves graph reconstruction compared to previous methods such as GRALE, since previous methods do not explicitly supervise node matching and graph reconstruction.

Critical strengths

Did not find any.

Critical weaknesses

The paper proposes a graph-level autoencoder using explicit supervision of spectral coordinates. The authors claim that their method improves graph reconstruction compared to previous methods such as GRALE, since previous methods do not explicitly supervise node matching and graph reconstruction. However, I have serious concerns regarding the claims, methodology, and experiments.

  1. The paper claims that GRALE's hidden representations are not directly supervised toward mathematically predefined node-level structural descriptors. However, GRALE already incorporates structural information into its node representations and jointly trains the encoder, decoder, and matcher through graph reconstruction loss. Therefore, the lack of explicit structural supervision does not necessarily mean that GRALE fails to learn meaningful structural information. To justify this claimed limitation and demonstrate the benefit of explicit structural descriptors, the authors should provide extensive ablation studies and controlled experiments showing that explicit supervision offers a clear advantage over the existing implicit structural learning approach.

  2. Sign ambiguity in Laplacian eigenvectors can lead to incorrect node matching, even when two nodes have the same structural information. Equation (5) directly compares the coordinates without checking whether a sign flip makes them equivalent. Equation (7) introduces SignAlign, but it is applied to the reconstruction loss, not the matching cost.

  3. Limited methodological novelty. The approach combines existing techniques and claims to design a new loss function. Also, this paper does not provide experimental ablation studies showing how the use of existing techniques improves performance or solves previous limitations. It remains unclear which component accounts for the reported improvements over GRALE. A slightly different loss function does not meet the standard of novelty.

    1. Using Hungarian matching to measure reconstruction accuracy is problematic (Appendix A.3). The main limitation of Hungarian matching is that it finds the best one-to-one assignment based on the given matching cost, not necessarily the best structural alignment between two graphs. Therefore, when the authors use this method in Equations (9) and (10), even if two graphs are isomorphic, an incorrect assignment can produce a nonzero graph edit distance.
  4. The predicted spectral coordinates are not guaranteed to produce a valid graph Laplacian, since the method does not enforce Laplacian properties such as zero row sums and non-positive off-diagonal entries.

  5. The evaluation focuses on three molecular reconstruction benchmarks with small graphs. There is insufficient evidence of robustness to graph symmetries, eigenvalue multiplicity, and node permutations.

  6. The paper claims a quadratic encoder complexity advantage over the cubic-complexity GRALE when spectral truncation is used. However, the experimental configurations in Table 5 use K=9,16,32K=9,16,32, matching the respective maximum graph sizes. Therefore, the author uses the full spectrum, for which standard dense eigendecomposition has cubic computational complexity. Thus, the claimed advantage does not hold for these configurations.

Questions for authors

N/A.

4 comments

a) The paper claims that GRALE's hidden representations are not directly supervised toward mathematically predefined node-level structural descriptors. However, GRALE already incorporates structural information into its node representations and jointly trains the encoder, decoder, and matcher through graph reconstruction loss. Therefore, the lack of explicit structural supervision does not necessarily mean that GRALE fails to learn meaningful structural information. To justify this claimed limitation and demonstrate the benefit of explicit structural descriptors, the authors should provide extensive ablation studies and controlled experiments showing that explicit supervision offers a clear advantage over the existing implicit structural learning approach.

We agree that GRALE incorporates structural information into its node representations, as also already highlighted in the paper and indeed we do not claim that GRALE fails to learn meaningful structural information. However, in our case we explicitly structure the model to learn three separate latent spaces capturing node, edge, and structural information, each with a corresponding loss term. This in turn allows us to use only node/edge feature or structure information for downstream tasks (see Table 2 of the original paper) and to interpolate only on some latent spaces (see Table 3 and Figure 5 of the original paper). The unseparated latent space design of GRALE does not allow for this. This is indeed one of the novel contributions of our approach. We will revise the manuscript to clarify this point better.

To be specific, on page 2 we state that GRALE’s matching representations are “not directly supervised to encode fine-grained node-level descriptors with explicit structural meaning.” Again, this does not imply that GRALE cannot learn meaningful structural information through reconstruction. Our main difference is that we explicitly supervise mathematically defined spectral coordinates and use the predicted coordinates jointly for graph decoding and node matching.

Indeed, the paper was missing an ablation analysis to show the importance of this structural supervision signal, as well as the other loss terms. As shown in the Tables included in the “Overall Response”, we have now performed an extensive set of ablation studies to verify this as well as other aspects of our model. Table 2 (overall response) shows the result of the loss terms ablation. This indicates the importance of the spectral/structural signals: removing the spectral losses causes the GED to increase from 0.0444 to 0.4060 and the GI Acc to drop from 97.44 to 65.63 on PubChem16. All these new results will be included in the appendix of the revised manuscript.

b) Sign ambiguity in Laplacian eigenvectors can lead to incorrect node matching, even when two nodes have the same structural information. Equation (5) directly compares the coordinates without checking whether a sign flip makes them equivalent. Equation (7) introduces SignAlign, but it is applied to the reconstruction loss, not the matching cost.

The reviewer is correct in observing that while GNNs guarantee permutation invariant, the eigendecomposition is not, i.e., node permutations can induce sign flips and random eigenbasis rotation in the computed eigenvectors. Our intention, however, was to allow the encoder to learn a stable representation wrt potential sign flips and bases rotations, while defining a matching strategy stable enough to not be impacted by a limited amount of sign flips and bases rotations. The robustness of the matcher stems from the observation that in general sign flips result in non-coherent signals for the matcher, while the coordinates with correct sign provide a coherent signal towards the correct match. For example, under the assumption of a 50% sign flip, half of the coordinates will effectively be noise, while the other will provide a useful signal towards finding the correct match. The signal is further reinforced by node feature information as well as information on the neighbourhood structure given by the probes (see Eq. 6 in the paper and Table 3 (overall response), which highlights the importance of the probe cost term).

Empirically, we observe that this robustness assumption holds true for PubChem16/32, as shown in Table 4 (overall response), where the performance on PubChem16/32 is not significantly impacted by random permutations of the test graphs (4th row), nor by random sign flips (2nd row) or a combination of this and random rotations (3rd row). Indeed, regardless of the perturbation the model still outperforms all the other approaches.

However, while performing these additional experiments, we have noticed that the same does not hold for QM9. We posit that this is due to the fact that in the training graphs of QM9 there is a consistent node ordering which provides a signal that our model is overfitting. We verify this hypothesis by re-training our model under different augmentation strategies and a new set of optimized hyper-parameters, i.e., (1) random node permutations, (2) random eigenvectors sign flips, and (3) random eigenvector sign flips + random orthogonal rotations of the eigenbasis in degenerate spaces. Table 5 (overall response) shows the results of this new experiment, highlighting that this simple training strategy allows our model to be robust wrt these perturbations.

In Table 8 (overall response) we perform an additional experiment where we measure the effect of introducing the sign correction in the matcher as well. This is achieved by alternating (a) the estimation of the permutation matrix P and (b) the estimation of the sign of the eigenvectors. The results show that this strategy has minimal impact on the reconstruction performance.

c) Limited methodological novelty. The approach combines existing techniques and claims to design a new loss function. Also, this paper does not provide experimental ablation studies showing how the use of existing techniques improves performance or solves previous limitations. It remains unclear which component accounts for the reported improvements over GRALE. A slightly different loss function does not meet the standard of novelty.

We would like to stress that the similarity with GRALE is largely confined to the use of a per-node representation which allows us to do unilateral node alignment (bipartite matching) instead of the bilateral alignment required by traditional graph matching. Our model is novel in several ways, not just the definition of a new loss term. As explained in response to previous comments, our architecture explicitly separates node, edge, and structural latent spaces by design. It also enables us in principle to use truncated spectral information, which in turn can potentially reduce the computational complexity at inference time. Also, the explicitly supervised spectral coordinates are part of the decoding path and provide a clearly defined structural target. The main distinction is that we reconstruct these coordinates from a structural latent and use them to guide both graph decoding and node matching, rather than relying solely on implicitly learned matching representations.

The new, extensive ablation experiments we have performed and reported in the “Overall response” shows the importance of the different architectural components of our model, particularly the spectral/structural supervision signal. We are indeed thankful to the reviewers for their comments, as a result of which we have performed a wide range of new experiments that we believe significantly strengthen the message of our paper.

d) Using Hungarian matching to measure reconstruction accuracy is problematic (Appendix A.3). The main limitation of Hungarian matching is that it finds the best one-to-one assignment based on the given matching cost, not necessarily the best structural alignment between two graphs. Therefore, when the authors use this method in Equations (9) and (10), even if two graphs are isomorphic, an incorrect assignment can produce a nonzero graph edit distance.

Note that we followed the same standard protocol used by GRALE. In general, computing the exact GED and GI Acc would be unfeasible for large graphs, which is why this approximation is used in practice. Specifically, we used the code provided by GRALE’s authors, i.e., class GraphAutoencoderMetric in https://github.com/KrzakalaPaul/GRALE/blob/main/GRALE/loss/objective.py. Note that using any sub-optimal match will only result in a reduction of the estimated GI Acc and an increase of the estimated GED, respectively. In other words, the matcher can be seen as providing a lower/upper bound for the GI Acc/GED results.

However, given the limited size of the graphs in QM9, we were able to compute the exact GED and GI Acc for these graphs, without relying on the model’s matcher and its predicted matching matrix. We used the VF2 algorithm to identify isomorphics graphs (GED=0) and NetworkX’s graph_edit_distance to measure the exact GED for the remaining graphs. Table 1 (overall response) shows that our method, when evaluated using the exact GED and GI Acc, still outperforms GRALE by a similar margin.

e) The predicted spectral coordinates are not guaranteed to produce a valid graph Laplacian, since the method does not enforce Laplacian properties such as zero row sums and non-positive off-diagonal entries.

The L_gram loss term (Eq. 7 in the paper) was indeed added to ensure that SS^T is the graph Laplacian (i.e., close to it in L1 sense). We observe that this loss term steadily decreases during training, while the ablation analysis of Table 2 (overall response) shows that this term does not appear to play an important role in the model’s performance. The latter can be explained by the fact that enforcing the reconstruction of the input eigenvectors is already a strong prior for the reconstruction of the input graph Laplacian.

f) The evaluation focuses on three molecular reconstruction benchmarks with small graphs. There is insufficient evidence of robustness to graph symmetries, eigenvalue multiplicity, and node permutations.

Note that we evaluated the reconstruction performance on the same datasets as GRALE, with the exception of COLORING, a synthetic dataset, which we swapped for the real-world dataset QM9, as in the former dataset the node features are largely unrelated to the graph structure. Note also that while GRALE performs a larger set of downstream task experiments, in all these experiments the autoencoder was still trained on PubChem32.

Table 6 (overall response) evaluates the robustness of our model wrt symmetries and eigenvalue multiplicity and compares it with that of GRALE. The results show that our model is indeed largely robust, particularly on the larger graphs of PubChem16 and PubChem32, where GRALE’s performance (especially on PubChem32) is negatively affected by the large number of symmetries and degenerate eigenspaces. We would also like to take this opportunity to stress that while we tried to replicate GRALE’s results (Table 1 of their paper) on PubChem16 and PubChem32 using their code and the recommended hyper-parameters configuration of both (a) their codebase and (b) their paper, the results we obtained were significantly lower (e.g., GED 0.277 and GI Acc 85.70% on PubChem16). In fairness to the authors of GRALE, we therefore reported their results (from Table 1 of GRALE’s paper) in our Table 1 (main paper). This in turn explains the discrepancy between the results of Table 1 (main paper) and the new ablation presented in Table 6 (overall response).

g) The paper claims a quadratic encoder complexity advantage over the cubic-complexity GRALE when spectral truncation is used. However, the experimental configurations in Table 5 use (K=9,16,32), matching the respective maximum graph sizes. Therefore, the author uses the full spectrum, for which standard dense eigendecomposition has cubic computational complexity. Thus, the claimed advantage does not hold for these configurations.

We would like to stress that the quadratic complexity of our model (with truncated spectrum) is not a main contribution of our work, as highlighted by the fact that this is only discussed briefly in the appendix. Also, note that the model configurations of Table 5 (appendix) refer to the models of Table 1 (main paper), which were trained using the full spectrum. We never claimed otherwise. However, we do have results using the truncated spectrum, i.e., Figure 3 (main paper), which shows the reconstruction performance on PubChem16 with truncated spectrum, and Tables 6 and 7 (appendix), which evaluate the performance of the encoder with truncated spectrum on downstream tasks.

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

This paper proposes G2SAE, a graph-level autoencoder that reconstructs attributed graphs through explicitly supervised Laplacian spectral coordinates. Its encoder produces separate graph-level structural, node-semantic, and edge-semantic latents. The decoder recovers spectral coordinates from the structural latent and uses their pairwise relations to condition graph reconstruction. A parameter-free matcher combines spectral, node-category, and neighborhood-probe costs to align predictions with targets. Experiments on QM9, PubChem16, and PubChem32 report improved reconstruction over GraphVAE, PIGVAE, and GRALE, with additional spectral truncation, representation probing, and molecular interpolation analyses.

Critical strengths
  1. Explicitly supervised spectral coordinates provide a meaningful intermediate representation between graph-level compression and graph reconstruction. Their pairwise relations directly condition connectivity and attribute prediction, giving the decoder a clear structural target beyond the final graph losses .

  2. The parameter-free matcher combines spectral position, node category, and neighborhood consistency. Its spectral probes refine correspondences without explicitly constructing dense edge-to-edge matching costs, providing a concrete way to use local structure when aligning reconstructed and target nodes.

  3. Reconstruction improves consistently across QM9, PubChem16, and PubChem32. On PubChem32, GI accuracy rises from 66.80% for GRALE to 82.07%, while GED falls from 0.78 to 0.57. The gains on larger molecules provide useful empirical support for the overall design (Table 1).

Critical weaknesses
  1. The contribution of each component remains unclear. Section 4.2 varies spectral truncation and latent dimensions, but does not separately ablate spectral supervision, spectral conditioning of the decoder, or neighborhood-probe refinement. Controlled ablations with comparable backbone and latent capacity would help establish which proposed component drives the gains.

  2. In Appendix A.3, GED and GI accuracy are evaluated under the matcher's Hungarian correspondence PHP_H, which is an approximation to the exact edit-distance optimal permutation. Consequently, the observed gains may reflect a combination of better graph decoding and more reliable node matching. Clarifying how the metrics behave under a matcher-independent attributed-graph isomorphism check would help readers better understand the respective contributions of these two components.

  3. Spectral invariance needs clarification. Equation (7) aligns signs only after matching, while the encoder and Eq. (5) depend on the eigenvector basis; repeated-eigenvalue rotations are unspecified. Table 3 also uses same-topology endpoints, which should share an invariant structural latent, yet fixing versus interpolating it yields different topology preservation rates (99.63% versus 76.53%). Clarifying preprocessing and testing permutation consistency would help explain this result.

Questions for authors
  1. Can you ablate the spectral coordinate and Gram losses, spectral conditioning of the decoder, and neighborhood-probe refinement separately, keeping the backbone and latent capacity comparable? Which component accounts for most of the reconstruction gain?

  2. Could you clarify how node matching was handled when evaluating the baselines in Table 1? Also, would it be possible to check how often a reconstructed graph is actually isomorphic to the target despite having a nonzero matched edit count under PHP_H (i.e., verifying GI accuracy with a standalone isomorphism test)?

  3. How are eigenvector signs and repeated-eigenvalue bases handled before encoding and matching? For the same-topology pairs in Table 3, why do the two interpolation protocols differ?

1 comment

a) Can you ablate the spectral coordinate and Gram losses, spectral conditioning of the decoder, and neighborhood-probe refinement separately, keeping the backbone and latent capacity comparable? Which component accounts for most of the reconstruction gain?

The results of this ablation are shown in Table 2 (see the “Overall Response”) and discussed also in reply to Reviewer ZHAK’s comments. Table 3 (overall response) shows the result of the matcher cost terms ablation, including the neighborhood-probe cost. The results show that all costs, including the probe, contribute to the performance of the overall model, with the spectral cost playing the largest role while the Gram loss is minimally impactful. Note that we incorporated this loss term to ensure the validity of the reconstructed Laplacian, as this term serves exactly that purpose. Note that ensuring the reconstruction of a valid Laplacian is also an explicit comment made by Reviewer 297q.

b) Could you clarify how node matching was handled when evaluating the baselines in Table 1? Also, would it be possible to check how often a reconstructed graph is actually isomorphic to the target despite having a nonzero matched edit count under (i.e., verifying GI accuracy with a standalone isomorphism test)?

We followed the same evaluation protocol as GRALE (using the author’s code) as explained in the reply to Reviewer ZHAK. We have also recomputed the values of the GI Acc and GED using a standalone exact isomorphism test and GED calculation. The results are shown in Table 1 (overall response) for QM9 (where graphs are small enough to be able to make these exact calculations) and show that our autoencoder still outperforms GRALE on this dataset by a similar margin to what shown in Table 1 of our paper.

c) How are eigenvector signs and repeated-eigenvalue bases handled before encoding and matching? For the same-topology pairs in Table 3, why do the two interpolation protocols differ?

Our aim was to allow the encoder to learn a stable representation wrt potential sign flips and bases rotations, while defining a matching strategy stable enough to not be impacted by a limited amount of sign flips and bases rotations. The robustness of the matcher followed from the observation that in general sign flips result in non-coherent signals for the matcher, while the coordinates with correct sign provide a coherent signal towards the correct match. For example, under the assumption of a 50% sign flip, half of the coordinates will effectively be noise, while the others will provide a useful signal towards finding the correct match. The signal is further reinforced by node feature information as well as information on the neighbourhood structure given by the probes (see Eq. 6 in the paper).

Empirically, we observe that this robustness assumption holds true for PubChem16/32, as shown in Table 4 (overall response), where the performance on PubChem16/32 is not significantly impacted by random permutations of the test graphs (4th row), nor by random sign flips (2nd row) or a combination of this and random rotations (3rd row). Indeed, regardless of the perturbation the model still outperforms all the other approaches.

However, while performing these additional experiments, we have noticed that the same does not hold for QM9. We posit that this is due to the fact that in the training graphs of QM9 there is a consistent node ordering which provides a signal that our model is overfitting. We verify this hypothesis by re-training our model under different augmentation strategies and a new set of optimized hyper-parameters, i.e., (1) random node permutations, (2) random eigenvectors sign flips, and (3) random eigenvector sign flips + random orthogonal rotations of the eigenbasis in degenerate spaces. Table 5 (overall response) shows the results of this new experiment, highlighting that this simple training strategy allows our model to be robust wrt these perturbations.

Regarding the interpolation results, note that Figure 5 (paper) uses two different molecules that share the same graph structure. In the first row we only interpolate the latent representations corresponding to node and edge features. In the second one, we also interpolate the structural representations. Our idea was to show that fixing one structural latent embedding helps preserve the topology while changing the semantic latents, supporting controlled molecular editing. However, this also suggests that the structural manifold is not robust wrt long-range interpolations. We will add a comment to this effect in the paper.

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

Graph-level autoencoders struggle to reconstruct graph structure and align nodes after compression. G2SAE predicts supervised Laplacian spectral coordinates and combines them with separate semantic latents for decoding and matching. On QM9, PubChem16, and PubChem32, it outperforms GraphVAE, PIGVAE, and GRALE in graph edit distance and exact reconstruction accuracy.

Critical strengths
  1. G2SAE outperforms all reported baselines across three datasets. On PubChem32, exact reconstruction accuracy improves from 66.80% to 82.07% over GRALE.
  2. The parameter-free matcher combines spectral coordinates, node categories, and neighborhood information, addressing structural inconsistencies in node alignment without an additional learned matching network.
Critical weaknesses
  1. The main reconstruction results do not report variation across random seeds. This makes it difficult to assess the stability of the reported improvements.
  2. The method handles eigenvector sign ambiguity, but does not clearly address basis ambiguity caused by repeated or close eigenvalues. This may affect node alignment.
  3. The paper claims that spectral supervision improves both reconstruction and node matching. However, it lacks ablations that isolate these effects. The current results do not establish which component drives the improvement.
Questions for authors
  1. How much improvement comes specifically from spectral supervision? Would the model still perform well without the spectral losses?
  2. How do training time and memory usage compare with GRALE, especially as graph size increases?
  3. GI accuracy is measured using the model's Hungarian alignment. Would independent graph isomorphism verification produce the same results?
1 comment

a) The main reconstruction results do not report variation across random seeds. This makes it difficult to assess the stability of the reported improvements.

We followed the standard experimental protocol of GRALE, using the code available at https://github.com/KrzakalaPaul/GRALE/. The results for PubChem16/32 are reported directly from Table 1 of GRALE’s paper (see also the note under Table 6 in “Overall Response”). QM9 however is a new dataset that we added to the experimental setup of our paper, so we measured the model mean GED and GI Acc (± std dev) over 3 seeds for our model, i.e., GED 0.0313 ± 0.0044, GI 99.313% ± 0.072.

b) How much improvement comes specifically from spectral supervision? Would the model still perform well without the spectral losses?

As shown in Table 2 (overall response), we performed an extensive set of experiments to evaluate the impact of the different loss terms. The result shows that removing the spectral losses causes the GED to increase from 0.0444 to 0.4060 and the GI Acc to drop from 97.44 to 65.63 on PubChem16. The change appears to be largely tied to L_coord, with L_gram having little impact on the reconstruction quality when L_coord is present. Note that we incorporated this loss term to ensure the validity of the reconstructed Laplacian, as this term serves exactly that purpose. Note that ensuring the reconstruction of a valid Laplacian is also an explicit comment made by Reviewer 297q.

c) How do training time and memory usage compare with GRALE, especially as graph size increases?

Table 7 (overall response) shows the per-epoch training time of G2SAE and GRALE on QM9, PubChem16, and PubChem32. The results underline the competitiveness of our approach, which achieves a lower or comparable per-epoch training time wrt GRALE. Note that on PubChem16 our model opts for a smaller batch size, increasing the number of iterations and thus the per-epoch training time. Setting the same batch size as GRALE reveals a lower training time, in line with the other datasets. Finally, Table 9 (overall response) reports the model parameters for our model and GRALE on the three datasets.

d) GI accuracy is measured using the model's Hungarian alignment. Would independent graph isomorphism verification produce the same results?

Note that we followed the same standard protocol used by GRALE. In general, computing the exact GED and GI Acc would be unfeasible for large graphs, which is why this approximation is used in practice. Specifically, we used the code provided by GRALE’s authors, i.e., class GraphAutoencoderMetric in https://github.com/KrzakalaPaul/GRALE/blob/main/GRALE/loss/objective.py. Note that using any sub-optimal match will only result in a reduction of the estimated GI Acc and an increase of the estimated GED, respectively. In other words, the matcher can be seen as providing a lower/upper bound for the GI Acc/GED results.

However, given the limited size of the graphs in QM9, we were able to compute the exact GED and GI Acc for these graphs, without relying on the model’s matcher and its predicted matching matrix. We used the VF2 algorithm to identify isomorphics graphs (GED=0) and NetworkX’s graph_edit_distance to measure the exact GED for the remaining graphs. Table 1 (overall response) shows that our method, when evaluated using the exact GED and GI Acc, still outperforms GRALE by a similar margin.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Overall Response

Overall Response

We thank the reviewers for their comments which allowed us to perform a number of additional experiments which we believe greatly strengthened our paper. We report and briefly comment on these results here, referencing them in the per-reviewer responses below. These results will then be incorporated in an updated version of the paper, submitted by the rebuttal deadline.

1. Exact GED and GI Accuracy

ModelGED (Hungarian)GI Acc (Hungarian)GED (Exact)GI Acc (Exact)
Ours0.03399.23%0.02299.28%
GRALE0.04697.91%0.03898.17%

Table 1. Exact GED and GI Acc computation on QM9. We use the VF2 algorithm to identify isomorphics graphs (GED=0) and NetworkX’s graph_edit_distance to measure the exact GED for the remaining graphs.

2. Loss terms ablation

GEDNode GEDEdge GEDGI Acc
Original0.05320.00880.044497.44
w/o L_Gram0.05790.00800.049997.50
w/o L_coord0.41470.05620.358585.66
w/o L_Gram & L_coord1.52380.40601.117865.63
w/o L_node9.55539.40400.15130.00
w/o L_adj11.60870.110611.49810.47
w/o L_edge5.95460.12845.82628.02

Table 2. Impact of the loss terms on the reconstruction performance (measured in terms of GED and GI Acc) on PubChem16.

3. Matching cost terms ablation

GEDNode GEDEdge GEDGI Acc
Original0.05320.00880.044497.44
w/o C^probe0.11230.01210.100296.23
w/o C^node0.12350.08230.041293.78
w/o C^spectral11.23000.019611.21041.73
C^spectral Only0.25490.12320.131789.83
C^node Only12.63960.004212.63541.42

Table 3. Impact of the matching cost terms on the reconstruction performance (measured in terms of GED and GI Acc) on PubChem16.PubChem16

4. Robustness wrt test graph permutations

Test graph perturbationQM9 GEDQM9 GIPubChem16 GEDPubChem16 GIPubChem32 GEDPubChem32 GI
None0.03399.230.05397.440.57482.07
Node permutation (recomputed spectrum)4.08825.450.13994.060.63780.58
Random eigenvector sign flips4.44118.750.15593.830.64480.58
Random basis rotation within degenerate eigenspaces0.61588.900.09695.730.62380.89

Table 4. Robustness wrt test graph permutations on QM9, PubChem16, and PubChem32.

Training graph perturbationGED/GI (no test graph perturbation)GED/GI (eigenvectors sign flips)GED/GI (eigenvectors sign flips + basis rotation)GED/GI (random node permutations)
None0.0432 / 99.175.9733 / 3.756.0076 / 3.545.4895 / 9.48
Random eigenvector sign flips0.0029 / 99.910.0044 / 99.880.0684 / 97.640.0584 / 98.13
Random eigenvector sign flips & basis rotation in degenerate eigenspaces0.0072 / 99.630.0107 / 99.590.0120 / 99.590.0137 / 99.56
Node permutation0.0185 / 99.500.1573 / 95.020.1646 / 94.840.0184 / 99.41

Table 5. Evaluation of the effect of different training set perturbation strategies on the robustness of our model wrt random test graphs node permutations (on QM9).

5. Robustness wrt symmetries and eigenvalue multiplicity

DatasetMethodStratumnGEDGI Acc
QM9Ours|Aut_c(G)| = 174030.019399.55
QM9Ours|Aut_c(G)| > 125970.071698.31
QM9Oursm_max ≤ 184190.020799.50
QM9Oursm_max > 115810.098097.79
QM9Oursall100000.032999.23
QM9GRALE|Aut_c(G)| = 174030.045398.01
QM9GRALE|Aut_c(G)| > 125970.047997.61
QM9GRALEm_max ≤ 184190.048597.84
QM9GRALEm_max > 115810.032698.29
QM9GRALEall100000.046097.91
PubChem16Ours|Aut_c(G)| = 142310.049297.71
PubChem16Ours|Aut_c(G)| > 157690.056297.24
PubChem16Oursm_max ≤ 156360.048397.71
PubChem16Oursm_max > 143640.059697.09
PubChem16Oursall100000.053297.44
PubChem16GRALE|Aut_c(G)| = 142310.169089.20
PubChem16GRALE|Aut_c(G)| > 157690.317384.49
PubChem16GRALEm_max ≤ 156360.225786.99
PubChem16GRALEm_max > 143640.291885.82
PubChem16GRALEall100000.254586.48
PubChem32Ours|Aut_c(G)| = 129010.475484.18
PubChem32Ours|Aut_c(G)| > 170990.617081.21
PubChem32Oursm_max ≤ 143010.444185.47
PubChem32Oursm_max > 156990.675479.51
PubChem32Oursall100000.575982.07
PubChem32GRALE|Aut_c(G)| = 129010.460285.18
PubChem32GRALE|Aut_c(G)| > 170991.781048.18
PubChem32GRALEm_max ≤ 143011.198164.17
PubChem32GRALEm_max > 156991.548554.94
PubChem32GRALEall100001.397858.91

Table 6. Stratified analysis of the performance of GRALE and our model for different levels of symmetries (|Aut_c(G)|) and maximum eigenvalues multiplicity (m_max) on QM9, PubChem16, and PubChem32.

A brief note on Table 6. We would like to take this opportunity to stress that while in our paper we tried to replicate GRALE’s results (Table 1 of their paper) on PubChem16 and PubChem32 using their code and the recommended hyper-parameters configuration of both (a) their codebase and (b) their paper, the results we obtained were significantly lower (e.g., GED 0.277 and GI Acc 85.70% on PubChem16). In fairness to the authors of GRALE, we therefore reported their results (from Table 1 of GRALE’s paper) in our Table 1 (main paper). This in turn explains the discrepancy between the results of Table 1 (main paper) and the new ablation presented in Table 6, as the latter is a novel result not taken from GRALE’s paper.

6. Per-epoch training times

QM9PubChem16PubChem32
Ours71s309s (178s*)515s
GRALE114s263s845s

Table 7. Per-epoch training times for our model and GRALE. While on QM9 and PubChem32 the batch sizes are the same, on PubChem16 our model opts for a smaller batch size (64, vs 128 for GRALE). Here the asterisk (*) denotes the per-epoch training time with batch size equal to that of GRALE.

7. Dealing with eigenvector signs

GEDGI Acc
Original0.053297.44
Alternating estimations of P and signs (2 iterations)0.061697.19
Alternating estimations of P and signs (3 iterations)0.056797.37

Table 8. Impact on the reconstruction accuracy (on PubChem16) of introducing sign correction in the matcher by alternating (a) estimation of P and (b) sign correction for 2 (3rd row) or 3 (4th row) iterations.

8. Number of model parameters

DatasetOur ModelGRALE
QM96.45M3.25M
PubChem164.48M3.33M
PubChem328.41M15.38M

Table 9. Number of learnable model parameters for our model and GRALE on QM9, PubChem16 and PubChem32.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.