acceptodds
Under review as a conference paper at ICLR 2027

COME: Complementary Reliability Monitoring with Expert Allocation for Multi-Source-Free Graph Domain Adaptation

Abstract

This paper studies the problem of multi-source-free graph domain adaptation, which aims to adapt multiple pretrained source models to target graphs without access to source graph data. Existing approaches usually combine pseudo-labeling techniques with domain alignment to learn from unlabeled target graphs. Despite the progress, their performance can degrade when source predictions are overconfident due to overconfident pseudo-labels from unaligned source models and thus potential error accumulation. To address this problem, we propose a novel approach named Complementary Reliability Monitoring with Expert Allocation (COME) for multi-source-free graph domain adaptation. The core of our COME extracts complementary structural and temporal stability signals for target supervision and expert allocation, i.e., structural stability and temporal stability. From the structural perspective, we feed perturbed views of the target graph into source models and compare their predictions to estimate structural stability. From the temporal perspective, we train a target-adaptive model and accumulate disagreement between target predictions and memory-based pseudo-labels that characterize temporal stability. These conflict scores are further incorporated into node states to stabilize pseudo-label supervision during target model adaptation. During inference, we leverage complementary stability signals as a routing function to dynamically allocate both frozen source and trainable target experts for reliable target predictions. Experiments on three benchmark families show that COME achieves the best mean accuracy in all six transfer settings.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Unsupervised Graph Domain Adaptation, Multi-Source-Free Domain Adaptation, Graph Neural Networks, Mixture of Experts, Node Classification

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_EA6qOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework called Complementary Reliability Monitoring with Expert Allocation (COME) for multi-source-free graph domain adaptation. The proposed framework extracts complementary structural and temporal stability signals for target adaptation. A reliability-aware expert allocation method is also proposed for target prediction. The experiments show the effectiveness of the proposed method.

Critical strengths
  1. The idea of aggregating multiple-source experts for individual target nodes is interesting and technically sound.
  2. The use of structural stability and temporal stability is reasonable.
Critical weaknesses
  1. In Table 2, the ablation studies show that the improvement is inconsistent across different experimental settings. Removing RETP leads to a performance drop of 12.44% on CSBM,0.22% on Twitch(R, P, F, ES→DE), and 0.11% on Citation (A, D→C). In Table 6 of the appendix, the results also show that the results on CSBM are more sensitive. The reason should be carefully analysed.
  2. The ablation shows that the improvement of SMEP is relatively small on the Twitch dataset and the Citation dataset. The effectiveness of this module is not well proven.
  3. The sensitivity study shows that the results decrease a lot when L>2. However, the reasons for such significant degradation are not well explained.
  4. The proposed method involves some heuristic designs, such as the target-adaptive score and the reliability of the source model. These designs raise concerns about their generalizability.
Questions for authors

Please refer to the Critical Weaknesses section.

1 comment
AuthorsOpenReview ↗

Response to Reviewer EA6q

We thank the reviewer for the careful review, constructive suggestions, and valuable questions about component contributions, parameter sensitivity, and generalizability. We have strengthened the corresponding experiments and explanations as detailed below.

W1 Differences in ablation gains and parameter sensitivity

Thank you for this valuable question. We further investigated the task-dependent contribution of RETP by comparing target predictions before and after expert routing, measuring corrected and introduced errors and the resulting accuracy changes. Routing corrects errors on 6.37% of CSBM-G4 nodes while introducing only 0.88% new errors, improving accuracy from 87.77% to 93.50%. Across all six transfers, routing increases average accuracy from 75.13% to 76.21%. CSBM shows greater sensitivity because changes in within-class and between-class connections directly affect neighborhood aggregation and source-model reliability, while stronger stable-state supervision can amplify incorrect pseudo-labels. Further analysis is provided in Section 4.2 and Appendix D.1, with additional results in Appendices D.5 and D.6.

MetricCSBM-G4DEENDBLPv7ACMv9Citationv1
Corrected Errors6.37 ± 1.331.74 ± 0.330.37 ± 0.350.17 ± 0.080.77 ± 0.170.10 ± 0.03
Induced Errors0.88 ± 0.251.49 ± 0.390.28 ± 0.310.08 ± 0.070.53 ± 0.120.04 ± 0.01
MethodCSBM-G4DEENDBLPv7ACMv9Citationv1Avg.
Target Model Alone87.77 ± 1.8868.00 ± 0.0759.95 ± 0.1578.69 ± 0.1573.62 ± 0.1482.76 ± 0.2775.13
Full COME93.50 ± 0.8068.24 ± 0.0760.08 ± 0.1878.77 ± 0.1973.86 ± 0.1282.82 ± 0.2776.21

W2 Limited SMEP gains on Twitch and Citation

Thank you for raising this valuable concern. We further evaluated SMEP across six transfer settings by collecting all target-node–source-model prediction pairs and using AUROC to measure how well source reliability distinguishes correct from incorrect predictions. The results show AUROCs of 67.51% on CSBM and 64.91%–66.30% on Citation, compared with 52.30%–54.23% on Twitch. SMEP also improves average accuracy from 73.88% to 76.21%, with gains of 8.21% on CSBM and 1.57% and 3.04% on DBLPv7 and ACMv9. Further results and analysis are provided in Table 3, Obs. 4, and Obs. 10 of Section 4.2.

SignalCSBM-G4DEENDBLPv7ACMv9Citationv1Avg.
Source Reliability67.51 ± 0.0952.30 ± 0.0654.23 ± 0.1665.87 ± 0.0866.30 ± 0.1864.91 ± 0.1361.85
Target Historical Risk74.23 ± 3.6860.12 ± 5.6755.21 ± 0.6069.75 ± 1.7760.43 ± 1.7869.80 ± 0.3764.92

W3 Performance degradation beyond two graph layers

Thank you for this helpful comment. We have expanded the explanations in the main text and appendix to clarify the performance degradation with deeper GCNs. As the number of layers increases, repeated neighborhood aggregation mixes information from increasingly distant nodes, including nodes from different classes. This leads to oversmoothing and weakens class discrimination. In contrast, two layers balance neighborhood information and class discrimination, achieving the highest accuracy across all six transfer settings. Further discussion is provided in Section 4.2 and Appendix D.1.

W4 Generalizability of the heuristic scores

We further evaluated the proposed scoring strategies across six transfer settings through controlled comparisons. With view-averaged source predictions fixed, reliability weighting improves average accuracy from 73.07% with equal weighting to 76.21%. Similarly, keeping source and target predictions unchanged, dynamic expert allocation improves average accuracy from 74.07% with fixed fusion to 76.21%. The largest gains occur on DE, where accuracy increases from 61.15% to 68.24%, and CSBM-G4, where it rises from 91.15% to 93.50%. Both strategies improve accuracy across all six transfers, demonstrating their effectiveness across different graph domains. Further results are provided in Section 4.2 and Appendix D.4.

Reviewer_WiMeOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2(+1)
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes COME for multi-source-free graph domain adaptation in node classification. It evaluates frozen source GNNs on edge-perturbed target graphs to estimate node-specific source reliability. It also accumulates disagreement between target predictions and memory-based pseudo-labels to characterize target-node states. These signals guide target-model supervision and the final allocation of source and target experts. Experiments compare COME with 22 baselines across six transfer settings on CSBM, Twitch, and Citation. The paper reports the highest mean accuracy in each setting, together with module ablations, sensitivity studies, and comparisons across GNN backbones.

Critical strengths
  1. Adapting several pretrained graph models without source data is a useful problem. The framework offers a concrete way to combine source predictions with a target-adaptive model at the node level.

  2. COME achieves the highest mean accuracy in all six evaluated transfers, with an average of 76.21% compared with 75.06% for GraphATA. Results over five runs and the module ablations provide initial evidence that the combined framework is useful.

Critical weaknesses
  1. The questions raised at lines 56 and 60 concern source-model reliability and reliable pseudo-label supervision. These are common concerns in source-free and multi-source domain adaptation. COME does use graph edges and neighborhood information, but the paper does not clearly explain what new insight its structural and temporal signals provide into graph-specific problems such as conditional structure shift, changes in homophily, or the resulting errors in neighborhood aggregation. Appendix C.2 discusses the relation to GraphATA, but the additional contribution appears to be mainly a combination of manually designed reliability scores, temporal weighting, and expert routing. The structural module probes frozen source models using perturbed graphs. This measures sensitivity to the chosen perturbations; it does not itself make the source models more stable or establish that their predictions are more accurate. A source model can remain confidently wrong across all views. The paper needs stronger motivation and evidence that the score in Eq. (1) identifies reliable source predictions better than confidence or entropy alone. Likewise, low disagreement with memory-based pseudo-labels can reflect persistent errors. The module-removal ablations do not fully resolve this issue: removing SMEP changes both probing and source aggregation, while removing TMAC also removes state-conditioned supervision. These comparisons do not isolate the value of structural probing or temporal smoothing. Theorems 3.1 and 3.2 do not close this gap. Theorem 3.1 establishes a Lipschitz perturbation bound and a sufficient margin condition for preserving the predicted class. Preserving a prediction does not establish its correctness, and the result does not justify the full reliability score. Theorem 3.2 gives the variance reduction of an exponential moving average under constant conditional moments and zero temporal covariance. These assumptions are not validated for a learning process with an evolving model and memory bank. Neither result analyzes the benefit of multiple sources, the restrictions imposed by unavailable source data, or improvement in target prediction error. The results describe component properties, but provide limited theoretical support for the proposed adaptation mechanism.

  2. The related-work discussion misses several relevant lines of GDA research. Directly relevant source-free examples include Rank and Align [2] and Robust Cross Supervision with Target Mining [9]. Relevant work on graph structure and distribution shift includes Structural Re-weighting [1], Sparse Causal Discovery with Generative Intervention [4], Disentangled Graph Spectral Domain Adaptation [5], and the more recent DiffGDA [10] and ADAlign [11]. Hierarchical graph contrastive domain adaptation [6] is also relevant to the multi-source setting. These omissions weaken the positioning of the structural challenges introduced in the second paragraph and make the claimed novelty difficult to assess. The discussion should explain which ideas remain applicable under source-free constraints, with empirical comparisons using compatible protocols. Pair-Align and A2GNN are already cited, but their relevance to the structural argument is not developed and they are absent from the reported comparisons. For the graph-specific motivation, homophily-agnostic structure reconstruction [12] and homophily discrepancy analysis [7] are relevant to structural shift, while attribute-driven GDA [8] shows why node-attribute discrepancy also matters. The discussion should explain how COME's reliability signals relate to these forms of shift. Since these methods use source graphs during adaptation, their relevance is primarily conceptual unless they are adapted and evaluated under the same source-free constraints. The claim of “Comprehensive Validation” is also stronger than the evidence supports. Evaluation covers three benchmark families and six transfers. Broader coverage could include the larger MAG benchmark and the structurally different Airport benchmark discussed in GDABench [3], as well as an appropriate ogbn-arxiv temporal-transfer setting. The Citationv1 graph used here is derived from MAG, but is distinct from the MAG benchmark in GDABench. Blog is another relevant benchmark, although its two-domain setup would require an explicit explanation of how it is used to evaluate a multi-source method. The concern is whether the findings generalize across graph sizes and types of structural shift, rather than the absence of any one mandatory dataset. Finally, the paper provides no computational complexity analysis or runtime and memory comparison. COME requires source-model evaluation across multiple graph views, a target-adaptive model, memory-based neighbor retrieval, and expert allocation. The cost of structural probing should be separated from adaptation and prediction costs, with scaling in the number of sources, perturbation views, nodes, and edges. This matters because improvements over GraphATA on the five real-world transfers range from 0.32 to 1.53 percentage points, so the accuracy–efficiency trade-off remains unclear.

[1] Liu et al. Structural Re-weighting Improves Graph Domain Adaptation. ICML, 2023.

[2] Luo et al. Rank and Align: Towards Effective Source-free Graph Domain Adaptation. IJCAI, 2024.

[3] Liu et al. Revisiting, Benchmarking and Understanding Unsupervised Graph Domain Adaptation. NeurIPS, 2024.

[4] Luo et al. Sparse Causal Discovery with Generative Intervention for Unsupervised Graph Domain Adaptation. ICML, 2025.

[5] Yang et al. Disentangled Graph Spectral Domain Adaptation. ICML, 2025.

[6] Lin, Zhou, and Shen. Hierarchical graph contrastive domain adaptation for multi-source cross-network node classification. Expert Systems with Applications, 2025.

[7] Fang et al. Homophily Enhanced Graph Domain Adaptation. ICML, 2025.

[8] Fang et al. On the Benefits of Attribute-Driven Graph Domain Adaptation. ICLR, 2025.

[9] Luo et al. Robust Cross Supervision With Target Mining for Source-Free Graph Domain Adaptation. IEEE Transactions on Knowledge and Data Engineering, 2026.

[10] Chen et al. Learning Structure-Semantic Evolution Trajectories for Graph Domain Adaptation. ICLR, 2026.

[11] Chen et al. Learning Adaptive Distribution Alignment with Neural Characteristic Function for Graph Domain Adaptation. ICLR, 2026.

[12] Fang et al. Graph Domain Adaptation via Homophily-Agnostic Reconstructing Structure. AAAI, 2026.

Questions for authors
  1. What graph-specific failure mode does COME address beyond GraphATA and conventional confidence-based weighting? Can comparisons with confidence-only reliability and unsmoothed disagreement, while keeping the backbone and supervision terms fixed, isolate the contribution of the proposed signals?

  2. Do the proposed reliability and conflict scores predict actual source accuracy and pseudo-label correctness better than confidence alone, particularly under different structural shifts? How sensitive is this relationship to the edge-addition and edge-removal policy?

  3. What do Theorems 3.1 and 3.2 establish about reliable multi-source selection or target adaptation performance? Is the zero-temporal-covariance assumption in Definition 3.2 supported by the observed adaptation trajectories?

  4. How does COME compare with the omitted source-free methods under the same protocol? Can the authors provide results on a representative larger or structurally different benchmark, together with runtime and memory comparisons against GraphATA?

7 comments
AuthorsOpenReview ↗

Response to Reviewer WiMe (1)

We sincerely thank the reviewer for the thorough review and constructive suggestions. The detailed questions have helped us substantially improve the rigor, clarity, and experimental validation of our work. We have strengthened the graph-specific motivation, added controlled experiments and efficiency comparisons, and expanded the theoretical and experimental discussions. We address the related concerns together in the following seven topics.

1. Graph-specific motivation and the relation to GraphATA

Graph domain adaptation must account for changes in neighborhood structure, which can make different source models reliable for different target nodes. We have expanded the discussion of these structural challenges and clarified the distinction between COME and GraphATA. GraphATA constructs node-specific convolution parameters using local target context, whereas COME evaluates source-model reliability through structural perturbations and monitors target predictions over time. We use these complementary signals to guide pseudo-label supervision and dynamically allocate source and target experts for individual nodes. Further details are provided in Section 2 and Appendix C.2.

2. Error discrimination and the perturbation policy

To evaluate error discrimination and complementarity, we measured source reliability across all target-node–source-model pairs and historical risk against incorrect target predictions. We also compared the highest-risk 25% of nodes identified by each signal against the same memory-bank pseudo-label errors. The two scores achieve average AUROCs of 61.85% and 64.92%, while historical risk identifies 31.01%–78.95% of errors missed by structural monitoring, with conditional AUROCs of 55.44%–84.62%. We further examined the number of perturbed views from 1 to 7 under fixed edge-removal and addition ratios of 0.10 and 0.05, achieving the highest average accuracy of 76.21% at K=5K=5. These results demonstrate the error-discrimination capability and complementarity of the proposed signals. Further details are provided in Table 3 and Figure 4 of Section 4.2, Appendix C.3, and Appendix D.1.

SignalCSBM-G4DEENDBLPv7ACMv9Citationv1Avg.
Source Reliability67.51 ± 0.0952.30 ± 0.0654.23 ± 0.1665.87 ± 0.0866.30 ± 0.1864.91 ± 0.1361.85
Target Historical Risk74.23 ± 3.6860.12 ± 5.6755.21 ± 0.6069.75 ± 1.7760.43 ± 1.7869.80 ± 0.3764.92
AuthorsOpenReview ↗

Response to Reviewer WiMe (2)

3. Controlled comparisons of weighting and temporal accumulation

Thank you for highlighting the importance of isolating the contributions of individual design choices. We conducted three controlled comparisons by keeping view-averaged source predictions fixed for reliability weighting, retaining identical pseudo-label losses for temporal accumulation, and using the same source and target predictions for dynamic expert allocation. These strategies improve average accuracy from 73.07%, 72.86%, and 74.07% to 76.21%, respectively, with gains across all six transfers. We also measured routing outcomes and found that routing corrects more errors than it introduces in every transfer, including corrections on 6.37% of CSBM-G4 nodes compared with 0.88% induced errors. These results demonstrate the individual benefits of reliability weighting, temporal accumulation, and dynamic allocation. Further details are provided in Section 4.2 and Appendices D.4–D.6.

MethodCSBM-G4DEENDBLPv7ACMv9Citationv1Avg.
Equal Weighting86.97 ± 7.2961.30 ± 12.3059.10 ± 0.8675.67 ± 2.0073.10 ± 0.2782.26 ± 0.1473.07
Reliability Weighting93.50 ± 0.8068.24 ± 0.0760.08 ± 0.1878.77 ± 0.1973.86 ± 0.1282.82 ± 0.2776.21
Conflict ScoreCSBM-G4DEENDBLPv7ACMv9Citationv1Avg.
Current Step85.64 ± 7.2861.30 ± 12.3059.10 ± 0.8575.73 ± 2.0573.11 ± 0.3582.27 ± 0.1172.86
Moving Average93.50 ± 0.8068.24 ± 0.0760.08 ± 0.1878.77 ± 0.1973.86 ± 0.1282.82 ± 0.2776.21
MethodCSBM-G4DEENDBLPv7ACMv9Citationv1Avg.
Fixed Fusion91.15 ± 0.1061.15 ± 0.0560.03 ± 0.1777.61 ± 0.2373.61 ± 0.1180.86 ± 0.1274.07
COME (Dynamic)93.50 ± 0.8068.24 ± 0.0760.08 ± 0.1878.77 ± 0.1973.86 ± 0.1282.82 ± 0.2776.21
MetricCSBM-G4DEENDBLPv7ACMv9Citationv1
Corrected Errors6.37 ± 1.331.74 ± 0.330.37 ± 0.350.17 ± 0.080.77 ± 0.170.10 ± 0.03
Induced Errors0.88 ± 0.251.49 ± 0.390.28 ± 0.310.08 ± 0.070.53 ± 0.120.04 ± 0.01

4. Scope of the theoretical results and the covariance assumption

We have clarified how the theoretical results support the structural and temporal monitoring mechanisms. *Theorem 3.1 bounds changes in source predictions under structural perturbations and establishes a sufficient condition for preserving predicted classes. Theorem 3.2 shows that temporal accumulation reduces disagreement variance under local stationarity and zero temporal covariance. Furthermore, Eq. (22) establishes that the influence of an isolated disagreement decreases geometrically over time without requiring the covariance assumption. Together with the reliability validation and controlled comparisons above, these results provide theoretical and empirical support for the proposed monitoring mechanisms. Further discussion is provided in Sections 3.2 and 3.3 and Appendix B.2.

AuthorsOpenReview ↗

Response to Reviewer WiMe (3)

5. Related work and baseline protocols

We expanded the related-work discussion to better connect structural shifts, graph homophily, attribute discrepancies, and source-free adaptation. We also added DGSDA and MHGCDA as source-needed baselines under the corresponding source-access protocol, increasing the total number of baselines from 22 to 24. Across the six transfer settings, COME achieves the highest accuracy and improves the average from 65.58% for DGSDA, 69.94% for MHGCDA, and 75.06% for GraphATA to 76.21%. These additional comparisons strengthen the experimental evaluation against recent graph adaptation methods. Further discussion and results are provided in Sections 2 and 4.2, Table 1, and Appendix C.2.

MethodCSBM-G4DEENDBLPv7ACMv9Citationv1Avg.
DGSDA78.14 ± 0.3349.89 ± 0.1245.73 ± 0.0173.30 ± 0.0269.56 ± 0.0376.88 ± 0.0465.58
MHGCDA92.81 ± 0.1963.58 ± 0.2645.64 ± 0.0173.19 ± 0.0566.25 ± 0.0578.19 ± 0.2069.94
GraphATA90.14 ± 0.3666.71 ± 0.3959.56 ± 0.1478.45 ± 0.8773.17 ± 0.5282.33 ± 0.7575.06
COME93.50 ± 0.8068.24 ± 0.0760.08 ± 0.1878.77 ± 0.1973.86 ± 0.1282.82 ± 0.2776.21

6. Benchmark coverage and the scope of generalization

We examined the suggested benchmarks and their compatibility with multi-source-free graph domain adaptation. Our experiments cover six transfer settings across synthetic, social, and citation graphs, supported by additional reliability and controlled comparisons. Extending this evaluation to Blog requires a multi-source protocol beyond its two available domains, while Airport requires node-feature construction. The MAG benchmark and temporal ogbn-arxiv also require dedicated multi-source transfer settings, and our MAG-derived Citationv1 differs from the MAG benchmark in GDABench. We have clarified these dataset characteristics and protocol requirements in Section 4.1 and Appendix C.1.

7. Complexity and the accuracy–efficiency trade-off with GraphATA

We added time and space complexity analysis and evaluated runtime, GPU memory, and CPU memory across six target domains. COME reduces computational overhead by caching source-model computations, reusing source predictions, and recursively updating historical disagreement. The results show that COME achieves the highest average accuracy of 76.21%, outperforming GraphATA at 75.06%, while maintaining a moderate runtime of 45.97 ms per epoch. COME also runs faster than SHOT, GraphCTA, CAiDA, and MHGCDA, reducing runtime and peak GPU memory by 57.6% and 21.7% compared with GraphCTA. These results demonstrate that COME delivers superior accuracy with moderate computational cost among the evaluated methods. Further details are provided in Section 4.2, Appendices D.7 and D.8, and Table 15.

MethodAvg. Accuracy (%)Runtime (ms/epoch)Peak GPU (MiB)Peak CPU (MiB)
SHOT71.3053.75528.951,823.18
GraphCTA71.37108.482,524.592,225.94
CAiDA72.5211,207.18611.242,879.87
MHGCDA69.9460.61625.602,115.10
GraphATA75.0615.681,587.342,078.79
COME76.2145.971,975.842,410.98
Reviewer_yV4hOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a complicated framework to combine multi-source GNNs to produce pseudo-labels on the target graph and train a target-adaptive GNN with dynamically adjusted pseudo-labels.

For the initial predictions from multiple source GNNs, the method combines them using reliability scores estimated from structural perturbations. During target adaptation, it uses a memory bank to store pseudo-label distributions and node embeddings, generating updated pseudo-labels by averaging the stored predictions of nearest neighbors in the embedding space. It also monitors temporal prediction disagreement to dynamically adjust the training losses.

Critical strengths
  1. The overall framework is carefully designed. The authors provide theorems establishing bounds on prediction changes under structural perturbations and variance reduction through temporal disagreement aggregation. These results are important for supporting reliable source-model selection and reducing fluctuations in pseudo-label supervision, further strengthening the motivation behind the proposed framework.

  2. The authors provide strong benchmarking results. It provides the a comprehensive benchmarking covering source-needed, no-adaptation, single-source-free, and multi-source-free methods with an average accuracy of 76.21%, outperforming the strongest multi-source-free baseline, GraphATA (75.06%).

Critical weaknesses
  1. One of my main concerns is about the cost. Since it involves a multi-source GNNs in the target GNN. Will the cost increase a lot?
Questions for authors

N/A.

1 comment
AuthorsOpenReview ↗

Response to Reviewer yV4h

We thank the reviewer for the constructive suggestion and valuable question about computational cost. The feedback helped us clarify the practical requirements of COME.

W1 Computational cost of multiple source models

To evaluate the computational overhead of using multiple source GNNs, we compare COME with four representative methods across six target domains. We measure per-epoch runtime and peak GPU and CPU memory for each method on every domain, together with average classification accuracy. COME achieves the highest average accuracy of 76.21%, improving over SHOT, GraphCTA, CAiDA, and MHGCDA by 6.89%, 6.78%, 5.09%, and 8.96%, respectively, while achieving the shortest runtime of 45.97 ms per epoch. We also reduce peak GPU memory by 21.7% compared with GraphCTA and peak CPU memory by 16.3% compared with CAiDA. These results demonstrate that COME improves prediction accuracy while maintaining competitive computational efficiency. The complexity analysis and detailed measurements are provided in Section 4.2 and Appendices D.7–D.8.

MethodAvg. Accuracy (%)Runtime (ms/epoch)Peak GPU (MiB)Peak CPU (MiB)
SHOT71.3053.75528.951,823.18
GraphCTA71.37108.482,524.592,225.94
CAiDA72.5211,207.18611.242,879.87
MHGCDA69.9460.61625.602,115.10
COME76.2145.971,975.842,410.98
Reviewer_qa1qOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

COME adapts several pretrained GNNs to an unlabeled target graph without access to the source graphs. It checks how much each source model’s predictions change when edges are perturbed. During adaptation, it also tracks how much the target model disagrees with the pseudo-labels from its memory bank. These scores are used to weight pseudo-label supervision and combine source and target predictions for each node. COME reports higher mean accuracy than GraphATA in all six transfer settings, averaging 76.21% versus 75.06%.

Critical strengths
  1. I like that the method makes two separate decisions for each node: how much to trust its pseudo-label during training and how much to rely on the source models at inference. I can see why the same source model might be useful for one node and less useful for another.

  2. COME has higher mean accuracy than GraphATA in all six settings. The biggest gain is on synthetic CSBM, but the results also improve on both Twitch transfers and all three Citation transfers.

Critical weaknesses
  1. I’m still unsure what the two scores tell us about which predictions are wrong. A source model can give the same wrong answer after several edge perturbations. The target model and its memory bank can also agree on a wrong pseudo-label. The two theorems explain why predictions may stay stable under perturbation and why averaging can reduce variance. I wanted to see whether either score actually flags errors, especially errors the other score misses.

  2. I had trouble telling which part of the method explains the gains in Table 2. Removing SMEP also removes averaging across perturbed views. Removing TMAC changes the pseudo-label supervision as well as removing the moving average. So even though the drop without TMAC is large, it’s hard to tell how much of it comes from using the history of disagreement. I had a similar question about the final routing. Would a fixed mixture of the same source and target predictions give similar results?

Questions for authors
  1. Among source predictions with similar confidence, are those that are more stable under edge perturbations more often correct? When the structural score is high, can temporal disagreement still distinguish correct from incorrect target pseudo-labels?

  2. If the view-averaged predictions from each source model are held fixed, how does the reliability weighting in Eq. (2) compare with giving all source models equal weight?

  3. With the same hard and soft pseudo-label losses, how does using the moving average compare with using only the current step’s disagreement? Could you report the target model’s accuracy before applying the final routing?

  4. With the same source and target predictions, how does the proposed routing compare with a fixed mixture? Compared with the target model alone, how often does routing correct a wrong prediction, and how often does it turn a correct one into an error?

  5. How were the hyperparameters and final checkpoints selected, and were target labels used for either decision? Were the GraphATA results in Table 1 taken from the original paper or obtained from a rerun using the same source checkpoints and seeds as COME? If paired runs are available, could you report the per-seed accuracy differences on the real-data transfers?

3 comments
AuthorsOpenReview ↗

Response to Reviewer qa1q (1)

We sincerely thank the reviewer for the thoughtful and constructive feedback, and for recognizing the value of node-wise supervision and expert allocation. The detailed questions on reliability validation, component contributions, and experimental protocols have been particularly insightful and helped us substantially strengthen the rigor and clarity of our work. We have added targeted experiments, refined our analyses, and clarified the experimental settings to address these concerns, as detailed below.

W1 Error discrimination and complementary error identification

Prediction stability alone does not guarantee correctness, so we directly evaluated the error discrimination and complementarity of both reliability signals. We used AUROC to assess whether source reliability distinguishes correct source predictions and historical risk identifies incorrect target predictions. Across six transfers, the two scores achieve average AUROCs of 61.85% and 64.92%. We further compared the highest-risk 25% of nodes flagged by each signal against the same memory-bank pseudo-label errors. Historical risk identifies 31.01%–78.95% of errors missed by structural monitoring and achieves conditional AUROCs of 55.44%–84.62% among unflagged nodes, demonstrating their complementary value. Further details are provided in Table 3 and Figure 4 of Section 4.2.

SignalCSBM-G4DEENDBLPv7ACMv9Citationv1Avg.
Source Reliability67.51 ± 0.0952.30 ± 0.0654.23 ± 0.1665.87 ± 0.0866.30 ± 0.1864.91 ± 0.1361.85
Target Historical Risk74.23 ± 3.6860.12 ± 5.6755.21 ± 0.6069.75 ± 1.7760.43 ± 1.7869.80 ± 0.3764.92

W2 Separating design contributions and evaluating fixed fusion

Thank you for highlighting this important issue in the ablation analysis. We conducted three controlled comparisons to isolate the contributions of reliability weighting, temporal accumulation, and dynamic expert allocation. We kept view-averaged source predictions fixed when comparing reliability and equal weighting, retained the same hard and soft pseudo-label losses when comparing moving-average and current-step disagreement, and used identical source and target predictions when comparing dynamic and fixed fusion. These three designs improve average accuracy from 73.07%, 72.86%, and 74.07% to 76.21%, with gains across all six transfers. We also clarified the combined contribution of temporal modeling and state-conditioned supervision in the TMAC ablation. Further details are provided in Obs. 5 and Table 4 of Section 4.2, Appendix D.4, and Appendix D.6.

MethodCSBM-G4DEENDBLPv7ACMv9Citationv1Avg.
Equal Weighting86.97 ± 7.2961.30 ± 12.3059.10 ± 0.8675.67 ± 2.0073.10 ± 0.2782.26 ± 0.1473.07
Reliability Weighting93.50 ± 0.8068.24 ± 0.0760.08 ± 0.1878.77 ± 0.1973.86 ± 0.1282.82 ± 0.2776.21
Conflict ScoreCSBM-G4DEENDBLPv7ACMv9Citationv1Avg.
Current Step85.64 ± 7.2861.30 ± 12.3059.10 ± 0.8575.73 ± 2.0573.11 ± 0.3582.27 ± 0.1172.86
Moving Average93.50 ± 0.8068.24 ± 0.0760.08 ± 0.1878.77 ± 0.1973.86 ± 0.1282.82 ± 0.2776.21
MethodCSBM-G4DEENDBLPv7ACMv9Citationv1Avg.
Fixed Fusion91.15 ± 0.1061.15 ± 0.0560.03 ± 0.1777.61 ± 0.2373.61 ± 0.1180.86 ± 0.1274.07
COME (Dynamic)93.50 ± 0.8068.24 ± 0.0760.08 ± 0.1878.77 ± 0.1973.86 ± 0.1282.82 ± 0.2776.21
AuthorsOpenReview ↗

Response to Reviewer qa1q (2)

Q1 Stability at similar confidence and conditional error discrimination

We further examined whether structural reliability distinguishes correct source predictions and whether historical risk can identify errors missed by structural monitoring. The combined source-reliability score achieves an average AUROC of 61.85% across six transfers. We also evaluated historical risk among nodes outside the highest-risk 25% under structural monitoring. Historical risk identifies 31.01%–78.95% of the missed pseudo-label errors and achieves conditional AUROCs of 55.44%–84.62%, demonstrating that temporal information provides complementary error detection even for structurally unflagged nodes. Further details are provided in Table 3 and Figure 4 of Section 4.2.

Q2 Reliability weighting versus equal weighting with fixed source predictions

To isolate the contribution of reliability weighting, we kept view-averaged source predictions fixed and replaced only the source weights with equal weights of 1/M1/M, while retaining all other settings. Reliability weighting improves accuracy across all six transfers, increasing the average from 73.07% to 76.21%, with the largest gains on CSBM-G4 from 86.97% to 93.50% and DE from 61.30% to 68.24%. These results show the benefit of node-specific reliability weighting beyond view averaging. Further details are provided in Table 11 of Appendix D.4.

Q3 Moving-average comparison and target accuracy before routing

To assess temporal accumulation, we replaced the moving average in Eq. (6) with current-step disagreement by setting ρ=0\rho=0, while retaining the same hard and soft pseudo-label losses. We also compared target-model accuracy before and after final expert routing. Moving-average disagreement improves average accuracy from 72.86% to 76.21% across six transfers. Final routing further improves average accuracy from 75.13% to 76.21%, with the largest gain on CSBM-G4 from 87.77% to 93.50%. These results demonstrate the benefits of temporal accumulation and expert routing. Further details are provided in Tables 13 and 14 of Appendix D.6.

MethodCSBM-G4DEENDBLPv7ACMv9Citationv1Avg.
Target Model Alone87.77 ± 1.8868.00 ± 0.0759.95 ± 0.1578.69 ± 0.1573.62 ± 0.1482.76 ± 0.2775.13
Full COME93.50 ± 0.8068.24 ± 0.0760.08 ± 0.1878.77 ± 0.1973.86 ± 0.1282.82 ± 0.2776.21

Q4 Fixed fusion, routing corrections, and induced errors

We compared dynamic expert allocation with fixed fusion while keeping source and target predictions unchanged, and further measured how often routing corrects or introduces errors relative to the target model alone. Dynamic allocation improves average accuracy from 74.07% to 76.21%, outperforming fixed fusion across all six transfers. The additional error analysis shows that routing corrects more errors than it introduces in every transfer. On CSBM-G4, routing corrects 6.37% of target nodes while introducing errors on only 0.88%, achieving a net correction of 5.49% of nodes. These results demonstrate the effectiveness of node-specific expert allocation in improving target predictions. Further details are provided in Section 4.2 and Appendix D.5.

MethodCSBM-G4DEENDBLPv7ACMv9Citationv1Avg.
Fixed Fusion91.15 ± 0.1061.15 ± 0.0560.03 ± 0.1777.61 ± 0.2373.61 ± 0.1180.86 ± 0.1274.07
COME (Dynamic)93.50 ± 0.8068.24 ± 0.0760.08 ± 0.1878.77 ± 0.1973.86 ± 0.1282.82 ± 0.2776.21
MetricCSBM-G4DEENDBLPv7ACMv9Citationv1
Corrected Errors6.37 ± 1.331.74 ± 0.330.37 ± 0.350.17 ± 0.080.77 ± 0.170.10 ± 0.03
Induced Errors0.88 ± 0.251.49 ± 0.390.28 ± 0.310.08 ± 0.070.53 ± 0.120.04 ± 0.01
AuthorsOpenReview ↗

Response to Reviewer qa1q (3)

Q5 Model selection and the source of GraphATA results

To clarify model selection, we used a shared configuration of L=2L=2, K=5K=5, λsta=1\lambda_{sta}=1, and λcon=0.05\lambda_{con}=0.05. We selected the best source checkpoints based on source validation performance and did not use target labels for hyperparameter or checkpoint selection during adaptation. Figure 4 uses peak-accuracy checkpoints only for retrospective error analysis. For GraphATA, we contacted its authors and obtained preprocessing code and archived processed data. Our reproduction yielded lower accuracy than the published results, so we adopted the stronger published results for the main comparison. COME outperforms the published GraphATA results across all six transfers, improving average accuracy from 75.06% to 76.21%. Further details are provided in Section 4.2 and Appendices C.2 and C.3.

MethodCSBM-G4DEENDBLPv7ACMv9Citationv1
Published GraphATA90.14 ± 0.3666.71 ± 0.3959.56 ± 0.1478.45 ± 0.8773.17 ± 0.5282.33 ± 0.75
Our Reproduction86.49 ± 3.0163.04 ± 1.2258.62 ± 0.5274.74 ± 0.2869.75 ± 1.9079.86 ± 0.59
COME93.50 ± 0.8068.24 ± 0.0760.08 ± 0.1878.77 ± 0.1973.86 ± 0.1282.82 ± 0.27

Comments on the paper

1 comment

Dear Reviewers,

Please review the authors' rebuttal and respond or engage in the discussion as appropriate.

Thank you for your time and effort.

Best regards,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.