acceptodds
Under review as a conference paper at ICLR 2027

Prototype-guided Topology Augmentation for Edge-Imbalanced Graph Domain Adaptation

Abstract

Graph Domain Adaptation (GDA) transfers knowledge from labeled source graphs to unlabeled target graphs, alleviating label scarcity in real-world graph data. Class imbalance is prevalent in graphs and can lead to biased knowledge transfer in GDA. However, existing methods for class-imbalanced GDA primarily focus on imbalanced node counts across source classes, while overlooking edge imbalance. Empirically, we observe pronounced intra-domain edge imbalance and substantial differences in edge distributions across domains. Moreover, we reveal that larger cross-domain edge-distribution discrepancies are associated with lower GDA performance. Inspired by these findings, we propose ProTA, a plug-and-play topology augmentation framework for edge-imbalanced GDA. ProTA constructs topology-aware source prototypes through class-preserving diffusion to guide candidate edge selection. Augmentation budgets are allocated according to intra-domain edge deficits and cross-domain edge-distribution discrepancies. The selected edges strengthen relational support for minority-class nodes in both domains, reducing intra-domain edge imbalance and cross-domain edge distribution discrepancies. Experiments on 44 transfer tasks across four benchmarks demonstrate that ProTA improves the performance of six representative GDA models.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Domain Adaptation, Edge Imbalance, Topology Augmentation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_kgZMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors study graph domain adaptation under edge imbalance. They argue that existing class-imbalanced GDA methods mainly focus on the imbalanced number of nodes across classes, while overlooking intra-domain edge imbalance and cross-domain edge-distribution discrepancies. To address this issue, they propose ProTA, a plug-and-play topology augmentation framework. ProTA constructs topology-aware source prototypes through diffusion over a class-preserving source graph, and uses them together with node attribute similarity to score candidate edges for minority-class nodes in both domains. The augmentation budget for each class pair is then allocated according to intra-domain edge deficits and cross-domain edge-distribution discrepancies, and the augmented graphs are directly fed into existing GDA backbones. Experiments on multiple GDA benchmarks verify the performance of the proposed ProTA.

Critical strengths
  • The paper considers edge imbalance in graph domain adaptation, which complements existing class-imbalanced GDA methods that mainly focus on the imbalanced number of nodes across classes.

  • ProTA is designed as a plug-and-play module without modifying the backbone architecture, and experiments on multiple benchmarks verify the performance of the proposed ProTA.

Critical weaknesses
  • In Section 4.2, the minority class is identified using pseudo labels produced by the GDA backbone. How to ensure the identified minority class in the target domain aligns with the real minority class? ProTA adds edges to nodes in different classes to reduce edge-distribution discrepancies. Does this augmentation change the original homophily or neighborhood structure of the graphs?
  • Several recently published GDA methods can be included in the experiment to evaluate the performance of ProTA, such as [1] [2] [3].
  • Typos and presentation issues:
    • "consisits" in Section 1;
    • "summerize" in Section 5.1;
    • "improvedas" and the incomplete sentence "across the six In a few cases" in Section 5.3;
    • "Tables ??–??" in Appendix E.1;
    • Appendix E.4 contains only the section title without any content;

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

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

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

Questions for authors

Please carefully address the weaknesses.

4 comments

Part 1/3

We sincerely thank Reviewer 3 for the constructive and insightful comments. In response, we have conducted additional experiments on target minority-class identification and graph homophily, expanded comparisons with recent GDA methods, and revised the manuscript to improve its presentation. Our point-by-point responses are provided below.

W1. Reliability of Target-Domain Minority-Class Identification

1. Reliability of target minority-class identification

We appreciate the reviewer's insightful concern regarding the reliability of minority-class identification in the unlabeled target domain. We agree that inaccurate pseudo-labels may introduce biases into the estimated target-class distributions and affect subsequent topology augmentation.

To mitigate this issue, ProTA does not identify target minority classes directly from all predicted labels. Instead, as described in Section 4.2 (Eqs. 7–8), we employ a dual-criterion reliability filter based on prediction confidence and prototype-similarity margins. Specifically, target nodes are selected only when their prediction confidence exceeds τ\tau and their similarity margin relative to competing source-class prototypes exceeds γ\gamma. Target minority classes are then identified based on the pseudo-label counts of these reliable nodes, reducing reliance on uncertain predictions.

To directly evaluate the reliability of target minority-class identification, we compare UDAGCN and UDAGCN+ProTA across six transfer tasks on MAG and Twitch. As shown in Table R1, UDAGCN+ProTA achieves minority-class identification accuracies of 72.6%–90.2%, consistently outperforming UDAGCN across all six tasks. The improvements range from 20.4 to 29.9 percentage points, with an average gain of 23.98 percentage points. These results demonstrate that incorporating ProTA substantially improves the accuracy of target minority-class identification, providing empirical support for the effectiveness of the proposed framework.

Table R1. Minority-class accuracy (%) with UDAGCN and UDAGCN+ProTA.

DatasetTransferUDAGCNUDAGCN+ProTA
MAGDE→FR54.0 ± 0.2475.0 ± 0.65
MAGFR→DE52.2 ± 0.8872.6 ± 0.45
TwitchDE→ES64.1 ± 0.2285.3 ± 0.15
TwitchES→DE60.3 ± 0.1690.2 ± 0.36
TwitchDE→RU61.3 ± 0.5688.2 ± 0.99
TwitchRU→DE64.7 ± 0.4589.2 ± 0.19

2. Effect on homophily and neighborhood structure.

To examine whether ProTA changes graph homophily, we measure edge homophily before and after augmentation on four Twitch transfer tasks with TDSS. As shown in Table R2, the mean changes are small: the absolute change is at most 0.251 percentage points for the source graphs and 0.277 percentage points for the target graphs. Thus, the overall edge homophily remains close to its original level in these tested settings.

ProTA improves GDA performance by adjusting underrepresented class-pair connections and accounting for cross-domain edge-distribution discrepancies. Since it adds edges, it necessarily changes the neighborhoods of nodes incident to those edges. Nevertheless, the small changes in overall homophily indicate that these local modifications do not substantially alter the proportion of same-class edges at the graph level in the tested settings.

Table R2. Edge homophily before and after ProTA augmentation with TDSS
on Twitch transfer tasks.

TaskSource before (%)Source after (%)Change (pp)Target before (%)Target after (%)Change (pp)
DE→ES63.223463.3149 ± 0.0200+0.0915 ± 0.020057.999158.2760 ± 0.0294+0.2769 ± 0.0294
ES→DE57.999158.2493 ± 0.2859+0.2502 ± 0.285963.223463.1964 ± 0.0093-0.0270 ± 0.0093
DE→RU63.223463.3234 ± 0.0100+0.1000 ± 0.010061.762861.7636 ± 0.0014+0.0008 ± 0.0014
RU→DE61.762861.8579 ± 0.0200+0.0951 ± 0.020063.223463.2230 ± 0.0006-0.0004 ± 0.0006

Part 2/3

W2. Comparisons with Recent GDA Methods

We thank the reviewer for suggesting comparisons with more recent GDA methods. We agree that including these methods provides a more comprehensive evaluation of ProTA.

To address this comment, we have extended our experiments to include DiffGDA [1] and HGDA [3]. Specifically, we evaluate both methods and integrate ProTA into their respective frameworks. For GAA [2], whose implementation is not publicly available, we additionally report its published results from DiffGDA [1] on the same Citation transfer tasks. We have incorporated these experimental results into Tables 1–3 of the revised manuscript, with the corresponding parameters settings summarized in Table 17 and 18. We conduct experiments on 12 transfer tasks across four datasets, with representative results presented in Table R3.

(1) Comparison with recent GDA methods. Table R3 presents the results of DiffGDA and HGDA on four representative Citation transfer tasks, together with the published results of GAA. These additional results extend our evaluation to the recent methods suggested by the reviewer. For example, DiffGDA+ProTA achieves 73.44%, 76.58%, and 77.78% Ma-F1 on A→D, D→A, and C→D, respectively, compared with the reported GAA results of 68.07%, 62.66%, and 71.40%.

(2) Generalizability to recent GDA architectures. We further evaluate whether ProTA can complement DiffGDA and HGDA. As shown in Table 1, integrating ProTA consistently improves DiffGDA’s Ma-F1 across the four Citation tasks, with gains ranging from 0.50 to 1.48 percentage points. ProTA also improves HGDA’s Ma-F1 on all four tasks, with gains ranging from 0.23 to 0.83 percentage points. These results provide additional evidence that ProTA can be integrated with GDA methods beyond the six backbones originally evaluated. The complete experimental results are presented in Tables 1–3 of the main manuscript.

Table R3. Comparison with recent GDA methods and their ProTA-enhanced variants on four Citation transfer tasks (Mean ± Std).

MethodA→D Mi-F1A→D Ma-F1D→A Mi-F1D→A Ma-F1C→D Mi-F1C→D Ma-F1D→C Mi-F1D→C Ma-F1
GAA [2], reported in [1]73.32 ± 0.6768.07 ± 1.0968.32 ± 0.4362.66 ± 0.9376.04 ± 0.3971.40 ± 0.1678.27 ± 0.3172.74 ± 0.76
DiffGDA75.79 ± 0.2372.92 ± 0.3374.48 ± 0.7975.10 ± 0.5278.52 ± 0.6277.27 ± 0.3777.44 ± 0.1072.63 ± 0.34
DiffGDA + ProTA76.29 ± 0.2073.44 ± 0.3675.26 ± 0.8676.58 ± 0.4279.03 ± 0.6177.78 ± 0.4078.33 ± 0.0373.87 ± 0.06
Δ+0.50+0.52+0.78+1.48+0.51+0.50+0.89+1.25
HGDA32.64 ± 0.5811.88 ± 0.3529.87 ± 0.3612.42 ± 0.0733.59 ± 0.8415.26 ± 0.4428.09 ± 0.1310.34 ± 0.23
HGDA + ProTA33.04 ± 0.4812.68 ± 0.3230.28 ± 0.5713.25 ± 0.4834.81 ± 0.0915.49 ± 0.7127.60 ± 0.9611.17 ± 0.03
Δ+0.40+0.80+0.41+0.83+1.22+0.23−0.49+0.83

W3. Typos and Presentation Issues

We thank the reviewer for carefully identifying these typos and presentation issues. We have corrected all the reported errors, resolved the unresolved table references in Appendix E.1, and completed the missing content in Appendix E.4. Additionally, we have carefully reviewed the entire manuscript and corrected other similar issues to improve its clarity, consistency, and overall presentation. We appreciate the reviewer's constructive suggestions, which have helped us improve the quality of the manuscript.

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

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

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

Part 3/3

Other Changes in Our Revision During the Rebuttal

To further strengthen the empirical validation of ProTA and clarify its motivation and applicability, we have conducted additional experiments and analyses and incorporated the corresponding results into the revised manuscript. Specifically, we have expanded comparisons with recent GDA methods and cross-network transfer settings, further investigated the distinction between node imbalance and edge imbalance, examined alternative edge-distribution strategies for budget allocation, and evaluated the robustness and applicability of ProTA under different structural conditions. For ease of review, we summarize the key additional changes below:

  • Expanded baseline comparisons. We have incorporated two additional GDA methods, DiffGDA and HGDA, and evaluated their performance before and after integrating ProTA. For GAA, whose implementation is not publicly available, we additionally report its published results on the same Citation transfer tasks for reference. The corresponding results are presented in Tables 1–3 of the main manuscript and Table 6 of the Appendix.

  • Expanded cross-network transfer evaluation. We have added 22 additional transfer settings for the six original GDA baselines (UDAGCN, TDSS, DGSDA, ADAlign, PALA, and ImGDA), extending the evaluation beyond the four settings considered in the original submission. We report the performance of each baseline and its ProTA-enhanced counterpart, providing broader empirical evidence of ProTA's effectiveness across diverse cross-network transfer scenarios. The additional results are provided in Appendix Tables 7–13.

  • Additional robustness evaluation under varying structural imbalance. We have conducted controlled experiments to evaluate six GDA backbones under five levels of intra-domain edge imbalance and cross-domain edge-distribution discrepancy, varying the two structural factors independently. We compare the original backbones with their ProTA-enhanced counterparts using both Micro-F1 and Macro-F1. The results are presented in Appendix Figure 2, providing further evidence of ProTA's robustness to varying degrees of structural imbalance.

  • Distinction between node imbalance and edge imbalance. We have added an analysis of the relationship between node-class distributions and edge-pair distributions using the Blog1 and Blog2 datasets. Although both datasets exhibit relatively balanced node-class distributions, their edge-pair distributions remain substantially non-uniform and exhibit distinct connectivity patterns. This analysis highlights that node-level balance does not necessarily imply balanced edge distributions. The corresponding results are presented in Appendix Figure 1.

  • Comparison of global and conditional edge distributions. We have further compared two edge-budget allocation strategies: ProTA, which uses conditional neighbor distributions, and ProTA*, which uses global class-pair edge distributions. Experiments on eight Citation and Twitch transfer tasks show that both strategies achieve comparable downstream classification performance, suggesting that both global and conditional edge distributions can effectively guide edge-budget allocation. The results are provided in Appendix Table 3.

  • Evaluation under balanced node-class distributions. We have further evaluated an extended variant of ProTA on node-balanced but edge-imbalanced graphs. By adapting augmentation-target selection from minority classes to underrepresented class-pair connections, the extended variant preserves the core mechanisms of prototype-guided edge scoring and discrepancy-aware budget allocation. Experiments with UDAGCN on four transfer tasks show consistent improvements in both Micro-F1 and Macro-F1, with average gains of 1.97 and 2.85 percentage points, respectively. The results are presented in Appendix Table 5.

After checking the rebuttal, I keep my initial score.

Reviewer_E37POpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work introduces a plug-and-play topology augmentation framework for edge-imbalanced graph domain adaptation. It constructs topology-aware source prototypes through class-preserving graph diffusion to score candidate edges, and then assigns augmentation budgets by accounting for intra-domain edge deficits and cross-domain edge distribution gaps. The proposed model consistently improves classification performance, reduces edge imbalance, and produces more discriminative node representations through data-level modification of graph topology.

Critical strengths
  1. It highlights the under-explored issue of edge imbalance in GDA, distinguishing between intra-domain edge imbalance and cross-domain edge distribution discrepancy.
  2. Its two-fold mechanism is well-designed: prototype-guided candidate edge scoring and discrepancy-aware budget allocation, both supported by ablation studies.
  3. The experimental evaluation is comprehensive, along with imbalance analysis, t-SNE visualization, and hyperparameter sensitivity analysis.
Critical weaknesses
  1. The related‑work section cites a large volume of recent papers, but brief one‑sentence summaries for a few cited works are overly compressed with minimal contextual contrast.
  2. Even though the method mitigates edge imbalance, it does not dissect how augmented edges propagate influence through GNN message‑passing.
  3. Hyperparameter tuning is task-specific across diverse transfer settings. The parameter sensitivity experiments did not yield any conclusions that could help reduce the cost of parameter searching when faced with a new dataset.
Questions for authors

The proposed method relies on target pseudo-labels; however, how to ensure the pseudo-labels are reliable in your proposed framework?

4 comments

Part 1/4

We sincerely thank Reviewer 2 for the thoughtful and constructive comments. The reviewer’s feedback has helped us improve the clarity of our methodological presentation and strengthen the experimental analysis. In response, we have expanded the contextual discussion of related work, further clarified how topology augmentation influences GNN message passing, provided additional analyses of robustness and hyperparameter selection, and elaborated on the reliability of target-domain pseudo-labels. Our point-by-point responses are provided below.

W1. Depth and Contextual Comparison of Related Work

We thank the reviewer for this constructive suggestion. We agree that some descriptions of related work were overly concise and did not sufficiently convey the methodological context or highlight the distinctions among existing approaches.

To address this concern, we have revised Section 2 (Related Work) in the revised manuscript. Specifically, (1) we have expanded the descriptions of representative methods in class-imbalanced graph learning, explicitly distinguishing class-rebalancing approaches, which adjust class distributions or learning contributions, from topology-aware approaches, which address imbalance through graph structural information. (2) We have strengthened the discussion of GDA methods by providing more contextualized comparisons of their respective research focuses and technical mechanisms, including structural alignment, conditional distribution shifts, and homophily discrepancies. (3) We have further clarified the differences between existing class-imbalanced GDA methods and ProTA. In particular, while existing methods primarily focus on node-level class imbalance, ProTA explicitly addresses intra-domain edge imbalance and cross-domain edge-distribution discrepancies through topology augmentation.

Additionally, we have expanded Appendix D to provide more comprehensive methodological comparisons and further contextualize the relationship between existing approaches and ProTA.

W2. Influence of Topology Augmentation on GNN Message Passing

We thank the reviewer for this insightful comment. We clarify how ProTA's augmented edges affect GNN message passing and provide additional analyses of minority-class message propagation and classification performance under varying structural imbalance.

1. Influence of Augmented Edges on Message Passing

ProTA identifies underrepresented class-pair connections through class-conditional edge-budget allocation and selects candidate edges using prototype-guided scoring. Once added, these edges expand the one-hop neighborhoods of their endpoint nodes, allowing them to receive messages from newly connected neighbors. Depending on the GNN aggregation and normalization scheme, the augmented edges may also change the relative contributions of both new and existing neighbors to the aggregated representations. In multi-layer GNNs, these changes can further propagate through subsequent layers, extending the influence of augmented edges to higher-order neighborhoods.

2. Quantitative Analysis of Message Propagation to Minority-Class Nodes

To examine the effect of topology augmentation on message propagation, we consider UDAGCN under five levels of cross-domain edge-distribution discrepancy. We compare the proportion of messages received by minority-class nodes before and after ProTA augmentation. As reported in Table R4, ProTA consistently increases the proportion of messages received by minority-class nodes across all five levels. Specifically, this proportion increases from 10.90%–23.60% before augmentation to 33.00%–35.00% after augmentation, corresponding to gains of 11.40–22.10 percentage points. Notably, as the cross-domain discrepancy increases from 0.01 to 0.14, the improvement becomes more pronounced, increasing from 11.40 to 22.10 percentage points. These results demonstrate that ProTA substantially increases the availability of neighborhood messages to minority-class nodes, particularly under larger cross-domain structural discrepancies.

Table R1. Proportion of messages received by minority-class nodes before and after topology augmentation under different cross-domain discrepancies.

LevelDstD_{st}Before augmentation (%)After augmentation (%)Gain (pp)
10.0123.6035.00+11.40
20.0220.7034.60+13.90
30.0515.9034.20+18.30
40.0913.6033.80+20.20
50.1410.9033.00+22.10

Part 2/4

3. Impact of Improved Minority-Class Message Propagation on Classification Performance

To examine whether improved message propagation to minority-class nodes translates into better classification performance, we evaluate six GDA backbones with and without ProTA across five levels of intra-domain imbalance and cross-domain discrepancy. As reported in Table R3, ProTA improves Micro-F1 in all 30 comparisons under both structural factors. For Macro-F1, improvements are observed in 29 out of 30 comparisons under intra-domain imbalance and 28 out of 30 comparisons under cross-domain discrepancy. The average Micro-F1 and Macro-F1 gains reach 3.77 and 4.24 percentage points under intra-domain imbalance, and 6.21 and 6.01 percentage points under cross-domain discrepancy, respectively.

Combined with the message propagation analysis in Table R2, these results provide complementary evidence that increasing the proportion of neighborhood messages received by minority-class nodes can benefit their representation learning and contribute to improved graph classification performance. In particular, the consistent performance gains across different structural imbalance levels support the effectiveness of ProTA in alleviating the adverse effects of insufficient minority-class neighborhood information.

Table R2. Performance improvements and sensitivity to structural imbalance across six backbones.

Structural factorMetricImproved settingsAverage gain (pp)Baseline range (pp)ProTA range (pp)
Intra-domain imbalanceMicro-F130/30+3.775.470.83
Intra-domain imbalanceMacro-F129/30+4.246.570.82
Cross-domain discrepancyMicro-F130/30+6.219.901.19
Cross-domain discrepancyMacro-F128/30+6.0112.620.92

W3. Hyperparameter Tuning and Search Cost

We thank the reviewer for this observation. We agree that the original parameter sensitivity analysis did not provide sufficient practical guidance for applying ProTA to new transfer tasks.

We have strengthened the analysis in Appendix E.4 by examining the Micro-F1 sensitivity of λ\lambda, γ\gamma, β\beta, and BB across Citation and Twitch tasks with TDSS and ADAlign. The results show that parameter sensitivity varies across transfer tasks and backbones. In particular, some Twitch tasks exhibit substantial changes in Micro-F1 when varying (\lambda), (\gamma), or (\beta), while other task–backbone combinations are relatively stable. This suggests that the search budget should be allocated according to the observed sensitivity for the new task, rather than using an equally extensive search for every parameter.

Based on these observations, we recommend a staged search strategy: start with a small set of previously tested configurations, conduct a coarse search to identify sensitive parameters, and allocate more trials to those parameters while using fewer candidate values for less sensitive ones. We also propose Bayesian optimization as an option to select promising configurations based on previous trials, potentially reducing search cost. This guidance has been added to Appendix E.4.

Part 3/4

Q1. Reliability of Target Pseudo-labels

We thank the reviewer for this question. Inaccurate target pseudo-labels may affect minority-class identification and topology augmentation. To mitigate this risk, ProTA selects reliable target nodes using both prediction confidence and a source-prototype similarity margin. We further evaluate the accuracy of the selected pseudo-labels through additional experiments.

1 Dual-filtering mechanism for reliable pseudo-label selection.

Before topology augmentation, we warm up the GDA backbone on the original source and target graphs to generate target pseudo-labels. To improve their reliability, ProTA retains a pseudo-label only when it satisfies two criteria: its prediction confidence is at least τ\tau, and its source-prototype similarity margin is at least γ\gamma.

The confidence criterion filters out uncertain predictions. The prototype-margin criterion further checks whether the target node is
sufficiently more similar to the source prototype of its predicted class than to the most similar alternative class prototype. This
additional check can filter out some high-confidence but prototype-inconsistent pseudo-labels. Only the retained pseudo-labels
are used to estimate target class counts and construct candidate neighbor sets for topology augmentation.

2. Empirical validation of target minority-class identification.

To evaluate the reliability of the selected pseudo-labels, we compare three settings: no filtering, confidence-only filtering, and confidence plus prototype-margin filtering. We conduct this comparison with TDSS and DGSDA on 12 Citation and Twitch transfer settings. Target ground-truth labels are used only for post-hoc evaluation.

As shown in Table R3, confidence plus prototype-margin filtering achieves pseudo-label accuracies of 70%–86% across the tested settings, compared with 40%–57% for confidence-only filtering. The additional prototype-margin criterion improves accuracy in all 12 settings, with gains of 22–34 percentage points. These results support the effectiveness of the dual-criterion mechanism (our method) in selecting more accurate pseudo-labels .

Table R3. Target pseudo-label accuracy (%) under different selection criteria.

BackboneDatasetTransferNo filteringConfidence onlyConfidence + prototype
TDSSCitationC→D234570
TDSSCitationD→C244373
TDSSTwitchDE→ES285680
TDSSTwitchES→DE204072
TDSSTwitchDE→RU234579
TDSSTwitchRU→DE305278
DGSDACitationC→D214274
DGSDACitationD→C285785
DGSDATwitchDE→ES274976
DGSDATwitchES→DE315786
DGSDATwitchDE→RU255375
DGSDATwitchRU→DE265577

Part 4/4

Other Changes in Our Revision During the Rebuttal

To provide a more comprehensive evaluation of ProTA and further clarify its motivation and methodological contributions, we have conducted additional experiments and analyses and incorporated the corresponding results into the revised manuscript. These additions extend the comparisons with existing GDA methods, broaden the evaluation across transfer settings, and further examine edge imbalance and the effectiveness of the proposed augmentation strategy. For ease of review, we summarize the key additional changes below:

  • Expanded baseline comparisons. We have incorporated two additional GDA methods, DiffGDA and HGDA, and evaluated their performance before and after integrating ProTA. For GAA, whose implementation is not publicly available, we additionally report its published results on the same Citation transfer tasks for reference. The corresponding results are presented in Tables 1–3 of the main manuscript and Table 6 of the Appendix.

  • Expanded cross-network transfer evaluation. We have added 22 additional transfer settings for the six original GDA baselines (UDAGCN, TDSS, DGSDA, ADAlign, PALA, and ImGDA), extending the evaluation beyond the four settings considered in the original submission. We report the performance of each baseline and its ProTA-enhanced counterpart, providing broader empirical evidence of ProTA's effectiveness across diverse cross-network transfer scenarios. The additional results are provided in Appendix Tables 7–13.

  • Distinction between node imbalance and edge imbalance. We have added an analysis of the relationship between node-class distributions and edge-pair distributions using the Blog1 and Blog2 datasets. Although both datasets exhibit relatively balanced node-class distributions, their edge-pair distributions remain substantially non-uniform and exhibit distinct connectivity patterns. This analysis highlights that node-level balance does not necessarily imply balanced edge distributions. The corresponding results are presented in Appendix Figure 1.

  • Comparison of global and conditional edge distributions. We have further compared two edge-budget allocation strategies: ProTA, which uses conditional neighbor distributions, and ProTA*, which uses global class-pair edge distributions. Experiments on eight Citation and Twitch transfer tasks show that both strategies achieve comparable downstream classification performance, suggesting that both global and conditional edge distributions can effectively guide edge-budget allocation. The results are provided in Appendix Table 3.

  • Evaluation under balanced node-class distributions. We have further evaluated an extended variant of ProTA on node-balanced but edge-imbalanced graphs. By adapting augmentation-target selection from minority classes to underrepresented class-pair connections, the extended variant preserves the core mechanisms of prototype-guided edge scoring and discrepancy-aware budget allocation. Experiments with UDAGCN on four transfer tasks show consistent improvements in both Micro-F1 and Macro-F1, with average gains of 1.97 and 2.85 percentage points, respectively. The results are presented in Appendix Table 5.

Reviewer_zNAaOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies unsupervised graph domain adaptation under imbalance in class-pair edge distributions. It distinguishes within-domain nonuniformity from differences between source and target edge distributions, using Jensen-Shannon divergence (JSD) as a diagnostic. The proposed ProTA framework constructs source class prototypes through diffusion over same-class edges, identifies minority classes using source labels and filtered target pseudo-labels, and adds edges according to prototype-based scores and class-specific budgets. These budgets combine within-domain deficits with cross-domain discrepancies. The augmented graphs are then used by existing GDA models without changing their architectures or alignment objectives. Experiments cover four benchmarks and six backbones, with performance comparisons, component ablations, and edge-distribution analyses.

Critical strengths
  1. The framework is modular and addresses a relevant practical problem. ProTA modifies the source and target graphs while retaining the underlying GDA architecture and objective. Its evaluation with six backbones, including two designed for class imbalance, provides useful evidence that the augmentation can complement different adaptation methods.

  2. Some reported gains are substantial, and the paper includes meaningful component comparisons. For example, UDAGCN gains 17.45 Macro-F1 percentage points on D→A, and DGSDA gains 8.61 Macro-F1 points on DE→RU. The candidate-ranking ablation keeps the candidate set and edge budget fixed, which helps assess the value of guided selection. The five-run ablation results provide evidence beyond a single favorable run, although uncertainty is not reported consistently across the main results.

Critical weaknesses
  1. The global edge distribution depends on both class sizes and connectivity, so its distinction from node/label imbalance needs clarification. A uniform class-pair distribution is also not necessarily desirable. Moreover, Section 3.2 measures global JSD, whereas Appendix B.5 adjusts conditional neighbor distributions; their uniform references are not equivalent.

  2. Although experiments are repeated five times, the main tables omit uncertainty, and Figure 2(b) does not quantify the correlation's strength or significance. The largest Micro-F1 gain occurs with β=0, leaving the contribution of explicit cross-domain correction unclear.

Questions for authors

The following requests concern the current claims and could change my assessment; they are not suggestions for unrelated future work.

  1. Could the authors report uncertainty for the F1 gains and JSD changes, and quantify the correlation in Figure 2(b), accounting for tasks that share graphs?

  2. What distinguishes edge imbalance from node/label imbalance, and why is a uniform edge distribution an appropriate reference? How does ProTA handle edge imbalance when node classes are balanced?

  3. How do the conditional budget rules in Appendix B.5 align with the global JSD measures? Could the authors separately establish the contribution of the cross-domain correction term?

8 comments

Part 1/6

We appreciate your detailed feedback. Below, we address the concerns and questions raised in the weaknesses section. Please feel free to reach out if further clarification is required.

W1. Edge Imbalance and Budget Allocation

We thank the reviewer for the comments on the relationship between edge imbalance and node/label imbalance, the use of uniform references, and the connection between global and class-conditional edge distributions. Below, we clarify these distinctions and provide empirical evidence supporting our budget allocation strategy.

1. Distinction between Edge Imbalance and Node/Label Imbalance

Distinction between node-label and edge distributions. We agree that the global edge distribution is influenced by both class sizes and connectivity. To clarify their distinction, we examine the Blog datasets. The node-class distributions are relatively balanced in both graphs. In Blog1, the largest class contains 1.31 times as many nodes as the smallest class, while this ratio is 1.13 in Blog2. Nevertheless, edge frequencies can differ substantially even between class pairs with similar endpoint class sizes. For example, in Blog1, class 4 has 1.18 times as many nodes as class 1, whereas the number of 0–4 edges is 4.62 times that of 0–1 edges. The corresponding ratios in Blog2 are 1.13 and 4.57. These observations demonstrate that node-label proportions alone do not fully characterize the distribution of connections across class pairs. We have clarified this distinction in Appendix E.1 of the revised manuscript.

Table R1. Node-label and edge-type distributions in the Blog datasets.

DatasetNode class rangeNode IR0–1 edges0–4 edges(0–4 edges)/ (0–1 edges)
Blog113.91%–18.22%1.311.26%5.84%4.62
Blog215.78%–17.78%1.131.39%6.34%4.57

2. Uniform class-pair edge distribution

We agree that a uniform global class-pair edge distribution is not necessarily desirable for GDA. For example, within-class connections naturally dominate in homophilous graphs, and enforcing uniformity could introduce unnecessary cross-class edges and disrupt informative neighborhood structures.

ProTA does not aim to enforce a uniform edge distribution. Instead, it seeks to identify and selectively supplement underrepresented class-conditional connections while accounting for cross-domain structural discrepancies. In GNN, node representations are learned by aggregating neighborhood information. Consequently, differences in class-conditional neighbor distributions may cause nodes of the same class to aggregate different types of information across domains, potentially hindering representation alignment and knowledge transfer. By selectively augmenting underrepresented connections and accounting for cross-domain discrepancies in edge-budget allocation, ProTA aims to provide more favorable conditions for cross-domain knowledge transfer.

Importantly, ProTA uses a uniform distribution as a reference for identifying relatively underrepresented neighbor classes, rather than imposing it as a target distribution. The added edges are constrained by a limited per-node budget and selected using prototype-guided scores; In target domain, they are further restricted to reliable nodes. Consequently, ProTA does not require either graph to become uniform or the two graphs to have identical edge distributions.

Part 2/6

3. Global and Conditional Edge Distributions.

In Section 3.2, we analyze global edge distributions and observe an association between cross-domain edge distribution discrepancy and GDA performance. This observation motivates us to mitigate edge imbalance through topology augmentation.

When designing the augmentation strategy, however, estimating target-domain edge distributions is challenging because ground-truth labels are unavailable and pseudo-label errors may introduce estimation bias. ProTA therefore first selects a reliable subset of target nodes based on prediction confidence and prototype-similarity margins. This allows ProTA to identify underrepresented connections for individual center classes and allocate the corresponding edge budgets more directly.

To examine whether global and conditional edge distributions can both guide edge-budget allocation, we compare ProTA with ProTA*, which uses global class-pair edge distributions for budget allocation. Table R2 reports results on eight Citation and Twitch transfer tasks using DGSDA, averaged over five random seeds. The two variants achieve comparable performance, with absolute differences ranging from 0.02 to 0.34 percentage points in Micro-F1 and from 0.01 to 0.50 percentage points in Macro-F1. The average absolute differences are only 0.18 and 0.29 percentage points, respectively. These results suggest that both global edge distributions and conditional neighbor distributions can serve as effective signals for edge-budget allocation, achieving comparable downstream classification performance.

Table R2. Comparison of conditional and global edge distributions for edge-budget allocation with DGSDA.

DatasetTaskProTA Micro-F1 (%)ProTA* Micro-F1 (%)ProTA Macro-F1 (%)ProTA* Macro-F1 (%)
CitationA→D77.57 ± 0.3177.36 ± 0.1575.50 ± 0.3975.17 ± 0.20
CitationD→A74.21 ± 0.2674.25 ± 0.2575.13 ± 0.1875.22 ± 0.31
CitationC→D78.65 ± 0.2378.84 ± 0.2776.95 ± 0.2277.15 ± 0.16
CitationD→C83.21 ± 0.1682.87 ± 0.1482.08 ± 0.1781.81 ± 0.13
TwitchDE→ES67.69 ± 0.8167.71 ± 0.5557.80 ± 0.5657.81 ± 0.72
TwitchES→DE44.43 ± 0.1444.58 ± 0.1740.54 ± 0.3941.03 ± 0.50
TwitchDE→RU72.57 ± 0.7672.89 ± 0.9849.49 ± 0.5849.90 ± 0.79
TwitchRU→DE44.26 ± 0.1744.39 ± 0.2340.27 ± 0.5340.77 ± 0.70

4. Distinction between the uniform references.

We agree that the uniform references for global class-pair edge distributions and conditional neighbor distributions are not equivalent, as they are defined over different distribution spaces and serve different purposes. Specifically, the global uniform reference uR\mathbf{u}_{\mathcal{R}} is defined over all class-pair edge types and serves as a reference for quantifying intra-domain edge imbalance through JSD in Section 3.2. In contrast, the class-conditional uniform reference vc\mathbf{v}_c is defined over neighbor classes given a fixed central class cc and is used in Appendix B.5 to identify underrepresented neighbor classes for edge-budget allocation. We have clarified these definitions and their respective roles in Sections 3.2 and Appendix B.5 of the revised manuscript.

Part 3/6

W2. Uncertainty, Correlation, and Cross-Domain Correction.

We thank the reviewer for these comments. We have added standard deviations to the main and supplementary experimental results, quantified the correlations in Figure 2(b), and conducted additional controlled experiments to examine the contribution of cross-domain correction.

1. Uncertainty and correlation analysis.

We have added standard deviations over five independent runs to Tables 1–3 in the main paper and Tables 1–7 in the Appendix.

To quantify the relationship in Figure 2(b), we computed Pearson correlation coefficients between cross-domain edge distribution discrepancy and Macro-F1 across the 30 directed Twitch transfer tasks. As shown in Table R3, both TDSS and ADAlign exhibit negative correlations, indicating that larger cross-domain edge distribution discrepancies are associated with lower Macro-F1. We have added these correlation coefficients to Figure 2(b) in the revised manuscript.

Table R3. Pearson correlation between cross-domain edge distribution discrepancy and Macro-F1 across 30 Twitch transfer tasks.

MethodPearson correlation coefficient (rr)
TDSS-0.449
ADAlign-0.581

2. Contribution of Cross-Domain Correction

We thank the reviewer for this observation. ProTA's topology augmentation involves three key parameters: γ\gamma, β\beta, and λ\lambda. Specifically, τ\tau and γ\gamma jointly determine the selection of reliable target nodes based on prediction confidence and prototype-similarity margins, respectively. The parameter β\beta adjusts class-pair edge budgets according to cross-domain edge-distribution discrepancies, while λ\lambda controls the contribution of source-prototype consistency when ranking candidate edges.

Importantly, β\beta is not the only mechanism through which ProTA incorporates cross-domain information. Even when β=0\beta=0 disables explicit cross-domain budget correction, the reliability filter still uses source-class prototypes to identify reliable target nodes, and prototype-guided candidate edge scoring still incorporates source-domain information through λ\lambda. Thus, ProTA retains these cross-domain information mechanisms and can still achieve substantial performance gains without explicit cross-domain budget correction.

To further isolate the contribution of explicit cross-domain correction, we conduct controlled comparisons on eight Citation and Twitch transfer tasks. For each task, we compare the original ProTA configuration with an ablated variant obtained by setting β=0\beta=0, while keeping all other hyperparameters and components unchanged. As shown in Table R4, incorporating explicit cross-domain correction consistently improves both Micro-F1 and Macro-F1 across all eight tasks, with gains ranging from 0.97 to 6.50 percentage points in Micro-F1 and from 1.11 to 3.87 percentage points in Macro-F1. These results demonstrate that explicit cross-domain budget correction provides additional performance benefits beyond the other cross-domain information mechanisms retained in ProTA.

Table R4. Comparison of ProTA configurations with and without explicit cross-domain correction.

DatasetTaskProTA Micro-F1 (%)w/o correction Micro-F1 (%)ProTA Macro-F1 (%)w/o correction Macro-F1 (%)
CitationA→D77.57 ± 0.3175.51 ± 0.2875.50 ± 0.3973.41 ± 0.37
CitationD→A74.21 ± 0.2673.24 ± 0.4575.13 ± 0.1873.57 ± 0.33
CitationC→D78.65 ± 0.2376.86 ± 0.2276.95 ± 0.2273.08 ± 0.22
CitationD→C83.21 ± 0.1680.98 ± 0.1682.08 ± 0.1780.97 ± 0.16
TwitchDE→ES67.69 ± 0.8161.19 ± 0.4257.80 ± 0.5654.87 ± 0.47
TwitchES→DE44.43 ± 0.1442.02 ± 0.2340.54 ± 0.3939.19 ± 0.46
TwitchDE→RU72.57 ± 0.7668.88 ± 0.6849.49 ± 0.5846.24 ± 0.29
TwitchRU→DE44.26 ± 0.1742.64 ± 0.2540.27 ± 0.5338.35 ± 0.65

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

Dear Area Chair and Reviewers,

We sincerely thank you for your time and effort in reviewing our manuscript and for your detailed and constructive feedback. We greatly appreciate your recognition of the following strengths of our work:

  • Relevant and underexplored research problem. ProTA addresses a practically relevant problem (Reviewer zNAa). It highlights the underexplored issue of edge imbalance in GDA, distinguishing between intra-domain edge imbalance and cross-domain edge-distribution discrepancy (Reviewer E37P), and complements existing class-imbalanced GDA methods that mainly focus on imbalanced node-class distributions (Reviewer kgZM).

  • Modular and plug-and-play framework. ProTA modifies source and target graphs without changing the underlying GDA architecture or objective. Its evaluation with six backbones, including two designed for class imbalance, provides evidence of its compatibility with different adaptation methods (Reviewer zNAa). Its plug-and-play design and evaluation across multiple benchmarks were also recognized (Reviewer kgZM).

  • Well-designed mechanisms and meaningful ablations. The prototype-guided candidate edge scoring and discrepancy-aware budget allocation mechanisms are well-designed and supported by ablation studies (Reviewer E37P). The controlled candidate-ranking ablation, which keeps the candidate set and edge budget fixed, helps isolate the contribution of guided selection, while five-run ablation results provide additional experimental evidence (Reviewer zNAa).

  • Substantial classification improvements. ProTA achieves notable performance gains, including a 17.45-percentage-point Macro-F1 improvement for UDAGCN on D→A and an 8.61-percentage-point improvement for DGSDA on DE→RU (Reviewer zNAa).

  • Comprehensive experimental evaluation. The experiments cover multiple benchmarks and include edge-imbalance analysis, t-SNE visualization, and hyperparameter sensitivity analysis (Reviewer E37P). The evaluation across multiple benchmarks also supports the effectiveness of ProTA (Reviewer kgZM).

We have addressed the reviewers' comments point by point and strengthened the revised manuscript through methodological clarifications and additional experimental analyses. The main updates are summarized below:

  • Clarifying edge imbalance and its distinction from node imbalance. We expanded the related work discussion and added an analysis of node-class and class-pair edge distributions (Reviewers zNAa; Appendix Figure 1, p. 22).

  • Validating edge-budget allocation and cross-domain correction. We clarified the relationship between global and conditional edge distributions, compared their effectiveness in guiding edge-budget allocation, and conducted controlled ablations to isolate the contribution of explicit cross-domain correction (Reviewer zNAa; Appendix Tables 3–4, p. 22).

  • Evaluating robustness and broader applicability. We analyzed the influence of topology augmentation on GNN message passing and conducted controlled experiments under five levels of intra-domain edge imbalance and cross-domain edge-distribution discrepancy. We further evaluated an extended ProTA mechanism on node-balanced but edge-imbalanced graphs, demonstrating its effectiveness beyond node-imbalanced settings (Reviewers E37P and zNAa; Appendix Figure 2 and Table 5, p. 23).

  • Expanding comparisons with recent GDA methods. We incorporated DiffGDA and HGDA and evaluated their ProTA-enhanced variants across 12 transfer tasks on four benchmarks. We also included the published results of GAA for comparison, providing further evidence of ProTA's applicability to recent GDA architectures (Reviewer kgZM; Main Tables 1–3, pp. 7–8; Appendix Table 6, p. 24).

  • Extending cross-network transfer evaluations. Beyond the reviewer-requested comparisons, we added 22 new transfer settings for the six original GDA backbones, further strengthening the empirical validation of ProTA. (Appendix Tables 7–13, pp. 24–27).

  • Strengthening statistical reliability and target-domain analyses. We added standard deviations to the classification results and conducted exact QAP permutation tests to assess the statistical significance of the association between edge-distribution discrepancy and GDA performance. We also provided additional analyses of target pseudo-label reliability, minority-class identification, and graph homophily after augmentation (Reviewers zNAa, E37P, and kgZM; Figure 2(b), p. 2; Main Tables 1–3, pp. 7–8; Section 4.2, p. 5).

We hope these clarifications and additional experiments assist the Area Chair and reviewers in assessing ProTA's contributions. Thank you for your consideration. We are happy to provide further clarification if needed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.