acceptodds
Under review as a conference paper at ICLR 2027

Model-to-Data Distillation for Graph Neural Networks

16 views

Abstract

Graph neural networks (GNNs) increasingly rely on sophisticated architectures and training procedures to achieve desirable properties such as high predictive performance, fairness, and robustness. However, these properties typically remain tied to the models that learn them, limiting their transferability to simpler models and downstream settings. We introduce model-to-data (M2D) distillation, a new distillation paradigm that transfers properties learned by a complex teacher into graph data, enabling simpler models to recover them through standard training. M2D distillation explicitly trades model complexity for data complexity by jointly learning augmented node features and graph structure that encode the teacher’s behavior. The resulting graph serves as a persistent medium for knowledge transfer and can be used with different downstream models. Across multiple graph learning benchmarks, we show that M2D distillation enables simple GNNs to approximate the behavior of substantially more sophisticated teachers, including fairness-aware GNNs, Graph Attention Networks, and Graph Transformers, while maintaining comparable predictive performance and transferring desirable properties of the teacher.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Distillation, GNN, Fairness, Augmentation

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_bC4VOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Model-to-Data (M2D) distillation, which transfers knowledge from complex GNN teachers into graph data rather than directly into student models. The method learns augmented node features and graph structures through teacher-guided distillation. The resulting graph can be reused to train lightweight models without further teacher access. Experiments investigate fairness transfer, predictive performance, and generalization across different GNN architectures.

Critical strengths

S1. Interesting research perspective. The paper introduces model-to-data distillation as an alternative to conventional knowledge distillation. Encoding teacher knowledge into reusable graph data is an interesting direction with potential applications.

S2. Broad evaluation of teacher models. The experiments consider fairness-aware GNNs, GATs, and Graph Transformers. This demonstrates the potential applicability of the framework to different teacher architectures and learning objectives.

S3. Cross-architecture transferability. The paper evaluates whether distilled graphs can benefit independently trained GCN, SGC, and MLP models. The results provide useful evidence that teacher knowledge can be transferred through graph data.

Critical weaknesses

W1. Limited novelty of the technical implementation.

Although model-to-data distillation provides an interesting perspective, the proposed implementation mainly combines existing graph augmentation and knowledge distillation techniques. Feature masking, graph structure learning, and KL-based distillation are well-established methods. The paper does not sufficiently distinguish its technical contributions from previous teacher-guided graph augmentation and graph condensation methods. A more detailed comparison with closely related methods is needed to establish the methodological novelty.

W2. Inconsistent improvements over standard knowledge distillation.

Table 1 shows that M2D does not consistently outperform conventional knowledge distillation. For example, on Cora with a GAT teacher, the standard distillation baseline achieves 79.27% accuracy, while the feature-based M2D variant achieves 82.52%. However, with a Graph Transformer teacher on Cornell, the best M2D result reaches only 58.79%, compared with 70.59% for the teacher. Furthermore, the adjacency-based variant frequently underperforms the feature-based variant. These results raise questions about the effectiveness of the proposed framework across different architectures.

W3. Limited evidence for model-agnostic knowledge transfer.

A central claim is that the distilled graph can transfer teacher knowledge to different downstream architectures. However, the main cross-architecture experiments in Tables 2 and 3 are limited to Cora and NBA. The evaluated downstream models are also restricted to GCN, SGC, and MLP. These experiments provide preliminary evidence but do not fully support the general model-agnostic claim. Additional evaluations across more datasets and substantially different architectures would strengthen this conclusion.

W4. Insufficient evaluation of computational efficiency and scalability.

The proposed method requires iterative optimization of node features and graph structure. The structure learner estimates pairwise relationships between nodes, which introduces quadratic complexity with respect to the number of nodes. Although the appendix discusses computational complexity, the paper lacks empirical comparisons of training time, memory consumption, and scalability against standard distillation methods. It remains unclear whether the benefits of reusing the augmented graph justify the additional computational cost.

W5. Limited validation of the theoretical claims.

The paper presents two theorems concerning fairness preservation and attention knowledge transfer. However, Theorem 1 depends on excess prediction losses that are not empirically quantified. Theorem 2 relies on local linearization and assumptions about the student solution. The experiments provide qualitative visualizations and correlations, but they do not directly evaluate the tightness of the theoretical bounds or the approximation errors. This weakens the connection between the theoretical analysis and the reported empirical findings.

Questions for authors

Q1. Novelty. What are the key technical differences between M2D and existing teacher-guided graph augmentation or graph condensation methods? Can the authors clarify which components are genuinely novel?

Q2. Knowledge Transfer. Table 1 shows that M2D does not consistently outperform standard knowledge distillation. Under what conditions does M2D provide clear advantages over conventional distillation?

Q3. Generalization. Tables 2 and 3 evaluate cross-architecture transfer mainly on Cora and NBA. Can the authors provide additional evidence that the distilled graphs generalize to substantially different student architectures and datasets?

Q4. Computational Efficiency. What are the training time and peak GPU memory costs of M2D compared with standard knowledge distillation? How does the method scale to large graphs given the quadratic complexity of pairwise structure learning?

Q5. Theoretical Validation. Can the authors quantify the approximation errors or assumptions involved in Theorems 1 and 2? How well do the theoretical predictions align with the empirical results?

2 comments
AuthorsOpenReview ↗

Response to reviewer bC4V

We thank the reviewer for recognizing M2D's contributions. We address all concerns below.

W1, Q1 : Novelty

We agree that feature masking, structure learning, and KL-based distillation are established techniques. M2D's novelty lies in using the transformed graph as a reusable medium for transferring teacher knowledge. Existing methods combine graph augmentation and knowledge distillation [1–5], but none explicitly learns teacher-guided augmentation. Augmentation is generally performed independently of teacher predictions, and the transferred knowledge ultimately resides in the student model. We have clarified these distinctions in line 137.

[1] Knowledge distillation improves graph structure augmentation for graph neural networks. NeurIPS’22.

[2] T2-gnn: Graph neural networks for graphs with incomplete features and structure via teacher-student distillation. AAAI’23.

[3] Nosmog: Learning noise-robust and structure-aware mlps on graphs. ICLR’23.

[4] Graph augmentation guided federated knowledge distillation for multisite functional MRI analysis. Pattern Recognition.

[5] A cold-start aware group recommendation method based on an augmented graph and knowledge distillation. Applied Soft Computing.

W2, Q2: Knowledge Transfer

The reviewer’s CORNELL example compares M2D with the teacher rather than KD. With a GT teacher, M2D achieves 58.79% accuracy versus 52.98% for KD. Similarly, on CORA, M2D achieves 82.52% versus 79.27% for KD (Table 1). These improvements demonstrate that M2D can transfer teacher knowledge more effectively than standard KD, even when a lightweight student does not match the teacher's performance. The effectiveness of feature- and adjacency-based distillation varies. Feature distillation often performs better for classification, joint transformations improve accuracy on CORNELL and TEXAS (Table 1), and structural transformations often improve fairness (Fig. 4).

W3, Q3 : Generalization

Beyond CORA and NBA (Tables 2–3), the original Appendix Tables 8–10 evaluate independent GCN, SGC, and MLP across six additional datasets.

We further evaluate independently trained GAMLP [6] on graphs distilled using a FairGNN teacher and GCN student. GAMLP consistently improves teacher fidelity and reduces DP and EQOP across SIMULATION, GERMAN, and NBA, with accuracy reductions of up to 1.42 percentage points on SIMULATION and GERMAN, while improving accuracy on NBA.

Fairness and Accuracy of Independently Trained GAMLP

DatasetMethodAccuracy ↑Fidelity ↑DP ↓EQOP ↓
SimulationVanilla97.43 ± 0.2296.12 ± 0.1740.41 ± 0.135.19 ± 0.20
G̃ (Feat.)96.98 ± 0.2097.97 ± 0.1835.31 ± 0.153.17 ± 0.17
G̃ (Adj.)96.12 ± 0.3097.31 ± 0.2134.57 ± 0.283.69 ± 0.24
G̃ (Both)96.01 ± 0.1597.99 ± 0.2534.01 ± 0.113.13 ± 0.34
GermanVanilla73.20 ± 0.2993.01 ± 0.178.48 ± 0.326.13 ± 0.31
G̃ (Feat.)72.17 ± 0.1095.07 ± 0.194.57 ± 0.183.01 ± 0.14
G̃ (Adj.)72.37 ± 0.1694.12 ± 0.225.19 ± 0.122.98 ± 0.19
G̃ (Both)72.19 ± 0.2194.79 ± 0.265.00 ± 0.152.13 ± 0.23
NBAVanilla69.59 ± 0.1785.12 ± 0.139.19 ± 0.1521.97 ± 0.13
G̃ (Feat.)71.19 ± 0.1490.14 ± 0.116.57 ± 0.1717.18 ± 0.31
G̃ (Adj.)71.98 ± 0.2391.28 ± 0.336.12 ± 0.2216.51 ± 0.16
G̃ (Both)71.38 ± 0.3091.02 ± 0.276.01 ± 0.1215.19 ± 0.24

[6] Graph attention multi-layer perceptron. KDD’22.

W4, Q4: Computational Efficiency

We report runtime, peak GPU memory, and storage. M2D is more efficient than GT, FairVGNN, and NIFTY. We also demonstrate improved scalability without substantial accuracy loss through selective edge learning (see response to RzcJ (W1)).

W5, Q5 : Theoretical Validation

Thm. 1 bounds the fairness difference between teacher and downstream models through distillation error and excess log losses based on an unknown optimal predictor. Although the unknown optimal predictor makes direct evaluation of bound tightness challenging, results in Sec. 4.3 and Appendix Tables 8–10 are consistent with this characterization.

Thm. 2 relates teacher attention and learned graph. It employs first-order approximation, an established technique for analyzing neural nets [7,8], while explicitly accounting for approximation residuals rather than assuming exact linearity. Fig. 6–7 support this relationship: feature-mask similarity is positively associated with teacher attention, while learned edge weights correlate with GAT attention on CORA. See Appendix for results on CORNELL and GT.

We have further clarified assumptions and experimental connections (lines 302, 312, and 422).

[7] Neural Tangent Kernel: Convergence and Generalization in Neural Networks. NeurIPS’18.

[8] Wide Neural Networks of Any Depth Evolve as Linear Models Under Gradient Descent. NeurIPS’19.

AuthorsOpenReview ↗

Rebuttal Follow up

We would like to follow up on our rebuttal and ask whether our responses have addressed your concerns. We would appreciate any additional feedback and would be happy to clarify any remaining questions. If you find our response satisfactory, we would be grateful if you could increase your score.

Reviewer_ozfUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes model-to-data (M2D) distillation, which transfers teacher behavior into modified graph data rather than directly into student parameters. The framework alternates between training a student model and learning feature masks and graph structure using a teacher–student distillation objective. The authors evaluate predictive performance and fairness transfer using fairness-aware GNNs, GAT, and Graphormer as teachers. Additional experiments train new downstream models on the frozen augmented graphs to examine their reusability.

Critical strengths

S1. Transferring teacher behavior into reusable graph data is an interesting and practically relevant research direction.

S2. Independent downstream training experiments show that the learned graphs can transfer teacher behavior across models.

S3. The experiments evaluate both predictive performance and fairness, with ablations on feature and structure augmentation.

Critical weaknesses

W1. The paper lacks a key experimental control to support the contribution of teacher guidance. The authors should compare against the same graph modification modules trained only with ground-truth labels. Without this comparison, it is difficult to determine whether the improvements come from teacher knowledge or graph preprocessing itself.

W2. The authors claim that M2D introduces only limited changes to the original graph. However, the structural regularizer in Equation (5) controls node degrees and edge-weight magnitudes, rather than directly penalizing changes to the original connections. Therefore, its role in preserving the original topology is not sufficiently justified.

W3. The computational benefits are not sufficiently evaluated. Although the downstream models are simpler, learning the graph requires quadratic pairwise computation and may increase graph density. The paper lacks runtime, memory, and storage comparisons to demonstrate that the overall process is more efficient.

W4. Some experimental results require clarification. On PHOTO, Table 5 reports teacher and MLP (Feat.) accuracies of 91.35% and 86.94%, respectively. If distillation accuracy measures prediction agreement on the same test nodes, it should be at least 78.29%, whereas Table 7 reports 76.12%. The authors should explain whether these results use different metric definitions, teachers, or evaluation sets.

Questions for authors

Q1. Can the authors compare against graph modification modules trained only with ground-truth labels to clarify the contribution of teacher guidance?

Q2. How does the structural regularizer preserve the original topology? Can the authors report how much the graph connections change after augmentation?

Q3. Can the authors report the runtime, memory usage, and storage costs of graph construction and downstream training, compared with the teacher and standard KD?

Q4. How is distillation accuracy calculated, and how can the PHOTO results discussed in W4 be reconciled?

Q5. In Algorithm 1, does the graph learner update backpropagate through the student training step? This detail is not clearly explained.

4 comments
AuthorsOpenReview ↗

Response to reviewer ozfU

We thank the reviewer for recognizing M2D's relevance and cross-model knowledge transfer. We appreciate the feedback and address all concerns below.

W1, Q1: Contribution of teacher guidance

To isolate teacher guidance, we optimize M2D's feature and structure learning modules using only the classification loss (without distillation).

Fairness and Accuracy under Label-Only Graph Transformations

DatasetMetricGCN (Vanilla)GCN (Feat.)GCN (Adj.)GCN (Both)
SimulationACC ↑96.58 ± 0.1494.58 ± 0.1495.02 ± 0.2296.02 ± 0.14
DP ↓42.42 ± 0.4140.42 ± 0.2140.01 ± 0.0941.02 ± 0.11
EQOP ↓4.21 ± 0.374.15 ± 0.164.96 ± 0.294.01 ± 0.23
GermanACC ↑74.06 ± 0.5175.09 ± 0.2974.01 ± 0.1174.23 ± 0.17
DP ↓9.91 ± 0.539.94 ± 0.2210.01 ± 0.2810.03 ± 0.23
EQOP ↓7.18 ± 0.406.98 ± 0.137.39 ± 0.237.50 ± 0.25
NBAACC ↑72.23 ± 0.4272.18 ± 0.2171.30 ± 0.1870.02 ± 0.21
DP ↓8.10 ± 0.519.51 ± 0.198.72 ± 0.139.33 ± 0.10
EQOP ↓22.19 ± 0.6221.97 ± 0.1821.31 ± 0.1222.37 ± 0.24

Label-only transformations yield limited and inconsistent fairness improvements. DP decreases by only 2.41 percentage points on SIMULATION, with no improvement on GERMAN or NBA. EQOP improvements are similarly inconsistent, highlighting the role of teacher guidance in fairness transfer.

Accuracy under Label-Only Graph Transformations

DatasetGCN (Vanilla)GCN (Feat.)GCN (Adj.)GCN (Both)MLP (Vanilla)MLP (Feat.)
Cora79.23 ± 0.2681.79 ± 0.2180.75 ± 0.4481.80 ± 0.3263.76 ± 0.1165.98 ± 0.43
Citeseer71.64 ± 0.1670.79 ± 0.2670.72 ± 0.3071.20 ± 0.1568.71 ± 0.1465.12 ± 0.14
Photo90.23 ± 0.2385.91 ± 0.4687.91 ± 0.1288.33 ± 0.3885.03 ± 0.0985.00 ± 0.19
Cornell53.61 ± 0.4143.01 ± 0.2750.53 ± 0.1552.68 ± 0.4666.34 ± 0.1864.12 ± 0.23
Texas62.36 ± 0.2052.68 ± 0.1852.68 ± 0.1852.68 ± 0.1875.27 ± 0.2071.45 ± 0.41

Though graph augmentation methods [1, 2] have demonstrated improvements in node classification, our label-only transformations improve accuracy only on CORA. M2D improves accuracy over vanilla GCN across all five datasets.

[1] "Data Augmentation for Graph Neural Networks," AAAI’21.

[2] "Iterative Deep Graph Learning for Graph Neural Networks: Better and Robust Node Embeddings," NeurIPS’20.

W2, Q2: Structural preservation

M2D regulates structural changes through the adjacency update and regularizer in Eq. (5): A~t=(1−γ)A+γ[(1−β)Wt+βs(X)].\tilde{\mathbf A}^{t}=(1-\gamma)\mathbf A+\gamma[(1-\beta)\mathbf W^t+\beta s(\mathbf X)]. The skip connection retains the original adjacency with weight (1−γ)(1-\gamma), while regularization controls node degrees and edge magnitudes. By the triangle inequality,

\leq \gamma\big[(1-\beta)||\mathbf W^t||_F +\beta||s(\mathbf X)||_F+||\mathbf A||_F\big].$$ Thus, $\gamma$ controls structural deviations, while regularization constrains edge magnitudes without explicitly enforcing topology preservation. Fig. 10 in the original Appendix quantifies these changes using Frobenius distance. We have clarified in line 268 that the adjacency update and structural regularization jointly regulate graph modifications, while Eq. (5) does not directly penalize deviations from the original adjacency. > W3, Q3: Computational efficiency and scalability We report runtime, peak GPU memory, and storage across classification and fairness datasets. M2D is more efficient than GT, FairVGNN, and NIFTY. We also demonstrate improved scalability through selective edge learning, reducing computational costs without substantial accuracy loss. See response to RzcJ (W1) for details. > W4, Q4: PHOTO fidelity discrepancy We identified four PHOTO runs with a GAT teacher and MLP student that used inconsistent splits. We corrected the evaluation to compare predictions on identical test nodes and obtained an MLP (Feat.) fidelity of 81.37% (see Tables 7 and 9). > Q5: Optimization in Algorithm 1 M2D approximately solves Eq. (1) through alternating first-order bilevel optimization [3–5]. We first update the student using classification loss with fixed graph learners. We then optimize the graph and feature learners using teacher-guided KL divergence and regularization, holding student parameters fixed. Gradients propagate through the student's forward pass but not its parameter-update step, avoiding unrolling and higher-order gradients. We have clarified this in line 1119. [3] "Darts: Differentiable architecture search." ICLR’19. [4] "Learning data augmentation with online bilevel optimization for image classification." CVF’21. [5] "On implicit bias in overparameterized bilevel optimization." ICML’22.
AuthorsOpenReview ↗

Rebuttal Follow up

We would like to follow up on our rebuttal and ask if our responses have addressed your concerns. We welcome any additional feedback and would be happy to answer further questions. If you are satisfied with our responses, we would appreciate it if you could consider increasing your score.

Thank you for your response. After considering your rebuttal, I have decided to keep my score unchanged.

Please let us know if all your concerns have been addressed (we have provided an itemized response). We hope we have clarified the results in Tables 5 and 7 (W4) and that the reviewer will consider at least updating the soundness score for our paper (currently assigned as 1). We believe that all the soundness concerns raised in the review have been addressed in our rebuttal.

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

The paper proposes a new paradigm called Model-to-Data (M2D) distillation, which does not directly distill a teacher model into a specific student model, but rather “distills” the teacher’s behavior into the training data (enriched node features and graph structure).

Critical strengths

1.With its fresh perspective and clear focus, this idea is very interesting.

2.The distillation output is an expanded graph rather than student weights; any architecture or future model can reproduce the teacher’s behavior through standard training without needing to access the teacher again, demonstrating clear practical value.

Critical weaknesses
  1. Scalability is limited, and structural learning requires pairwise scoring.

  2. Performance drops significantly when the teacher is weaker than the student (the student in an MLP on a heterogeneous graph), but since the method itself lacks a mechanism for evaluating teacher quality or selective distillation, it is difficult to guarantee performance.

Questions for authors

please refer to the weakness.

3 comments
AuthorsOpenReview ↗

Response to reviewer RzcJ

We thank the reviewer for recognizing M2D's fresh perspective and value of reusable graph data for independent downstream training. We appreciate the constructive feedback and have addressed all concerns, as detailed below.

W1. Scalability:

We acknowledge that the current structure learner incurs quadratic complexity due to dense pairwise edge scoring. However, this is a limitation of the structure learning implementation rather than an inherent weakness of M2D. Similar quadratic complexity arises in existing graph structure learners and Graph Transformers [1–3].

We report runtime (s), peak GPU memory (MB), and graph storage (MB), comparing M2D with GAT, Graph Transformer (GT), standard knowledge distillation (KD), and fairness-aware GNNs. Runtimes include training and evaluation, with M2D additionally optimizing graph transformations.

Computational comparison with GAT, GT, and standard KD.

DatasetM2D RuntimeM2D MemoryM2D StorageGAT RuntimeGAT MemoryGAT StorageGT RuntimeGT MemoryGT StorageKD RuntimeKD MemoryKD Storage
Cornell2.9369.591.1930.3769.321.3232.87122.031.320.3069.011.32
Texas3.1568.511.320.4564.261.1933.05102.031.190.4663.981.19
Cora7.631422.2542.781.181445.1814.93255.167475.6414.930.871231.0614.93
Citeseer12.842804.3989.221.571399.2147.11399.0410181.8347.110.5261550.4547.11
Photo27.146225.82244.994.592957.8724.501798.2149443.0624.501.493157.2324.50

Computational comparison with fairness-aware GNNs and standard KD.

ModelSim RuntimeSim MemorySim StorageGerman RuntimeGerman MemoryGerman StorageNBA RuntimeNBA MemoryNBA Storage
M2D (Both)22.6572.983.8622.7073.733.9118.1840.700.77
FairGNN10.1972.442.3611.5272.450.146.4326.030.40
FairVGNN68.351332.802.3657.211116.860.1459.581122.940.40
NIFTY36.851192.942.3628.691186.080.1430.051111.190.40
Standard KD5.6380.592.365.4580.630.143.4177.940.40

M2D requires substantially less runtime and GPU memory than GT across all five datasets. Similarly, it is consistently faster and reduces peak GPU memory by approximately 93–97% compared with FairVGNN and NIFTY. Although GAT, FairGNN, and standard KD are faster, M2D produces reusable graph data that enables independent downstream training without further teacher access.

Improving scalability through selective edge learning: We restrict structure learning to existing edges, following selective edge learning strategies [4,5], without modifying the distillation objective. Compared with the full M2D costs reported above, this reduces runtime and GPU memory while improving accuracy on homophilic datasets. On heterophilic datasets, unrestricted learning achieves higher accuracy, suggesting the benefit of learning additional connections.

Scalability of M2D: Full Pairwise Structure Learning vs. Selective Edge Learning

DatasetFull M2D Acc.Existing Edges Acc.RuntimeMemory
Cora82.0582.976.491302.03
Citeseer72.0273.018.871784.01
Photo91.2791.3226.911508.54
Cornell54.1353.582.6667.19
Texas64.4962.393.0766.11

These results demonstrate that quadratic pairwise scoring is not an inherent requirement of M2D, and selective edge learning offers a practical direction for improving scalability while preserving the framework's core contribution: reusable teacher-guided graph transformations.

[1] Liu et al., "Towards unsupervised deep graph structure learning." WWW, 2022.

[2] Franceschi et al., "Learning discrete structures for graph neural networks." ICML, 2019.

[3] Ying et al., "Do transformers really perform badly for graph representation?" NeuRIPS, 2021.

[4] Liu et al., "Towards Unsupervised Deep Graph Structure Learning," WWW 2022.

[5] Zheng et al., "Robust Graph Representation Learning via Neural Sparsification," ICML 2020.

W2. Teacher Quality and Negative Transfer:

We agree that teacher quality influences knowledge transfer. M2D aims to preserve teacher-learned behavior rather than guarantee improvements across all downstream architectures. When the teacher performs worse than the student, distillation may introduce negative transfer, particularly on heterophilic graphs where MLPs can outperform certain GNN teachers. We have clarified that M2D is most beneficial when the teacher offers desirable properties, such as improved accuracy or fairness, that downstream models cannot easily recover independently (line 1286).

Thank you to the author for the response, which has made this paper much clearer.

AuthorsOpenReview ↗

Thank you.

Thank you for taking the time to review our response and provide your thoughtful feedback. We appreciate your constructive comments and are glad that our response helped clarify the paper. If our response has addressed your concerns, we would be grateful if you could consider increasing your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.