acceptodds
Under review as a conference paper at ICLR 2027

MIRAGE: Mutual Information-guided Reconstruction and Adaptation for Graph Domain Adaptation with Edge Incompleteness

Abstract

Graph Domain Adaptation (GDA) aims to transfer knowledge from a labeled source graph to an unlabeled target graph. Existing GDA methods typically rely on fully observed graph topologies for representation learning and cross-domain alignment. However, real-world graph topologies can be incomplete, with edges unobserved due to corruption, adversarial perturbations, or privacy constraints. In this paper, we formulate this scenario as *Edge-Incomplete Graph Domain Adaptation* (EIGDA), where domain shift coexists with structural incompleteness in source and target graphs. Moreover, we show that EIGDA poses a key challenge: missing edges compromise available neighborhood information and can further exacerbate the source-target performance gap. To address this issue, we propose **MIRAGE** (**M**utual **I**nformation-guided **R**econstruction and **A**daptation for **G**raph domain adaptation with **E**dge incompleteness), a novel algorithm that leverages mutual information maximization to guide transfer-aware edge selection, thereby allowing graph reconstruction to preserve local information while supporting cross-domain adaptation. Experiments on multiple benchmark datasets with varying missing-edge ratios show that MIRAGE consistently outperforms competitive baselines under structurally incomplete EIGDA settings.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Domain Adaptation, Graph Neural Network, Domain Adaptation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_mWwqOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies graph domain adaptation when both source and target graphs have missing edges. The authors call this setting Edge-Incomplete Graph Domain Adaptation (EIGDA). Using GRADE, they show that the domain gap grows as more edges are removed. Their method, MIRAGE, keeps the observed edges and considers adding edges between nodes with similar features, using top-K neighbors. It predicts a probability for each candidate edge and adds the edge if this probability exceeds a threshold computed from the two node embeddings. Training combines a binary cross-entropy loss on the observed adjacency, mutual information (MI) objectives within and across domains, and source classification on the reconstructed graph.

Critical strengths

1.This paper makes sense in considering domain alignment when reconstructing the graph. Table 5 also suggests that applying VGAE before GDA is not enough. This is the main reason I lean toward acceptance. 2.The paper evaluates four datasets and five deletion rates. It also includes BFS-based deletion, additional transfer directions, ablations, and analyses of hyperparameters, efficiency, and variance. The gains at 60–80% deletion are large. I would find them more convincing if the concerns below were addressed.

Critical weaknesses

1.UDAGCN reports the same accuracy at every deletion rate on Blog and Airport: 0.167 and 0.249, respectively. These results suggest that it may be predicting a single class. On Blog, six of the eight GDA baselines also perform worse than GCN+VGAE, which does not use domain adaptation. 2.The paper does not report which edges MIRAGE adds. Specifically, BCE loss treats all unobserved candidate edges as negatives, while the threshold increases with the embedding dot product. More evidence is needed to show that reconstructing edges with domain alignment in mind is what drives the gains.

Questions for authors

1.How were the baselines tuned? In particular, why does UDAGCN report identical accuracies across deletion rates on Blog and Airport?Do the baseline results at 0% deletion match previously reported results? 2.Could you evaluate StruRW and HGDA with the same top-K edges? If they use degrees from the full graph, they retain information about edges deleted later. Could this explain why MIRAGE's accuracy changes so little on Airport? 3.How many edges are added, and what fraction connect nodes with the same label? Could the authors also compare the dynamic threshold with a fixed threshold? 4.Could you include StruRW and HGDA in Table 4?

5 comments
AuthorsOpenReview ↗

Response to Reviewer mWwq

Part 1/3

We thank the reviewer for the detailed feedback and constructive suggestions. We address the concerns and questions below and provide additional experimental results where requested.

1. Baseline Results and Experimental Settings (W1, Q1)

  • UDAGCN reports the same accuracy at every deletion rate on Blog and Airport: 0.167 and 0.249, respectively. These results suggest that it may be predicting a single class

    How were the baselines tuned? In particular, why does UDAGCN report identical accuracies across deletion rates on Blog and Airport?Do the baseline results at 0% deletion match previously reported results?

    Thank you for this question. For all baselines, we use the default hyperparameters provided in their publicly available implementations. For UDAGCN, these include 200 training epochs and an encoder dimension of 16, with a random seed of 200.

    Regarding the identical UDAGCN results across deletion rates, we examined its predictions and confirmed that all target nodes are predicted as the same class on both Blog and Airport. This explains why its accuracy and Macro-F1 remain unchanged across different deletion rates.

    Regarding the results at 0% deletion, the datasets used in our experiments differ from those in the original UDAGCN paper, which does not report results on Blog or Airport. Therefore, a direct numerical comparison with its originally reported results is not possible. We evaluate all baselines on the same datasets and transfer tasks for a consistent comparison.

  • On Blog, six of the eight GDA baselines also perform worse than GCN+VGAE, which does not use domain adaptation.

    We agree with this observation. Unlike the GDA baselines that directly perform adaptation on the input graphs, GCN+VGAE first reconstructs the source and target graphs using VGAE before node classification. This reconstruction may provide useful neighborhood information for classification, even without explicit domain adaptation. In contrast, missing edges can affect the node representations learned by conventional GDA methods and make cross-domain alignment more difficult.

2. Comparison with StruRW and HGDA (Q2)

We conduct additional experiments with two baselines, kkNN-Aug+StruRW and kkNN-Aug+HGDA, as reported in Table R1 (also added as Table 7, in lines 979-981 on page 19, highlighted in blue). Both methods preserve the observed edges and add edges based on node feature similarity, with the number of added edges matched to MIRAGE separately for each domain. The results show that MIRAGE consistently outperforms the strongest baseline by 6.4–14.6 percentage points across all six settings, demonstrating that feature-based edge augmentation alone cannot explain MIRAGE's improvements.

MethodB1→B2 60%B1→B2 80%A→D 60%A→D 80%B→U 60%B→U 80%
kkNN-Aug+StruRW0.3710.348‾\underline{0.348}0.5140.4870.3250.307
kkNN-Aug+HGDA0.375‾\underline{0.375}0.3430.554‾\underline{0.554}0.523‾\underline{0.523}0.408‾\underline{0.408}0.396‾\underline{0.396}
MIRAGE0.5100.4940.6180.5960.5480.539

Table R1. Comparison with kkNN-Aug+StruRW and kkNN-Aug+HGDA. The best results are in bold, and the second-best are underlined‾\underline{underlined}.

Regarding the concern about information leakage, all methods, including MIRAGE, StruRW, and HGDA, use only the observed incomplete graphs during training. Any degree information is computed from the graphs after edge deletion, without access to the original complete adjacency matrices.

On Airport, MIRAGE's accuracy decreases by only 0.019 from 0% to 80% edge deletion, compared with 0.074 and 0.077 for StruRW and HGDA, respectively. This suggests that MIRAGE is less sensitive to missing edges, consistent with its MI-guided reconstruction, which selects additional connections to mitigate domain shift rather than relying only on the observed graph structures.

AuthorsOpenReview ↗

Response to Reviewer mWwq

Part 2/3

3. Reconstructed Edges and Their Role in Domain Adaptation (W2, Q3)

The paper does not report which edges MIRAGE adds. Specifically, BCE loss treats all unobserved candidate edges as negatives, while the threshold increases with the embedding dot product. More evidence is needed to show that reconstructing edges with domain alignment in mind is what drives the gains.

How many edges are added, and what fraction connect nodes with the same label?

We thank the reviewer for these questions. First, we clarify that Eq. (2) is not optimized independently. As shown in Eq. (5), MIRAGE jointly optimizes adjacency reconstruction and MI objectives. In particular, structure-level cross-domain MI depends on candidate-edge probabilities and provides additional gradients during optimization. Thus, although BCE penalizes unobserved candidate edges, these edges can still be selected through MI-guided reconstruction. The threshold based on embedding similarity does not independently determine which edges are selected, as the learned embeddings and candidate-edge probabilities are influenced by the joint objectives.

To characterize the reconstructed edges, we report the numbers of added edges and their same-label proportions on Citation and Airport under 60% and 80% edge deletion rates in Table R2.

DatasetDeletionDomainAdded EdgesSame-label (%)
Citation60%ACMv9390483.1
DBLPv7285084.6
Citation80%ACMv9597182.3
DBLPv7442083.9
Airport60%Brazil30561.3
USA392174.6
Airport80%Brazil41660.1
USA534574.5

Table R2. Number of added edges and percentage connecting nodes with the same label.

The results show that MIRAGE adds more edges as the deletion rate increases, with a substantial proportion connecting nodes of the same class. We also examine the effect of reconstruction on source-target homophily discrepancy in Section 7 (in lines 523-529 on page 10, highlighted in blue), where MIRAGE achieves smaller homophily gaps than both the observed graphs and feature-kkNN augmentation.

More importantly, MIRAGE aims to reconstruct connections that address domain discrepancy rather than recover the original topology exactly. To examine this distinction, we compare MIRAGE with its variant without cross-domain MI on ACMv9→DBLPv7. As shown in Table R3, removing cross-domain MI improves edge recovery F1 but reduces target classification accuracy.

VariantRecovery F1 60%Target Acc. 60%Recovery F1 80%Target Acc. 80%
MIRAGE0.3330.6180.3060.596
MIRAGE w/o cross-domain MI0.4600.5520.4010.535

Table R3. Comparison of edge recovery F1 and target classification accuracy on ACMv9→DBLPv7.

Removing cross-domain MI improves edge recovery F1 but decreases target accuracy by 0.066 and 0.061 at 60% and 80% deletion, respectively. This shows that a higher overlap with the original topology does not necessarily lead to better domain adaptation.

We provide further evidence in Appendix G (Figs. 8 and 9, on page 20, highlighted in blue). The edge overlap analysis shows that MI components influence which edges are selected, while the classification loss gap analysis shows that full MIRAGE consistently achieves smaller source-target performance discrepancies than its MI ablation variants. At 80% deletion, full MIRAGE reduces the losWs gap by 13.6% on ACMv9$\to$DBLPv7 and 25.6% on Brazil$\to$USA compared with the variant without MI. These results support the role of MI-guided reconstruction in improving domain adaptation under edge incompleteness.

AuthorsOpenReview ↗

Response to Reviewer mWwq

Part 3/3

4. Dynamic vs. Fixed Thresholds (Q3)

We additionally compare the dynamic threshold with three fixed thresholds (0.4, 0.6, and 0.8) on ACMv9$\to$DBLPv7 (A$\to$D) and Blog1$\to$Blog2 (B1$\to$B2) under 60% and 80% edge deletion rates. The results are reported in Table 10 of Appendix I on page 22 . The dynamic threshold achieves the highest target classification accuracy in all four settings, outperforming the best fixed threshold by 0.012–0.024. These results demonstrate the effectiveness of dynamically adjusting the threshold based on node representations for edge selection.

5. Additional Results under BFS-Based Edge Deletion (Q4)

We have added StruRW and HGDA to the BFS-based edge deletion experiment, with the new results highlighted in blue in lines 439-441 on page 9. MIRAGE consistently outperforms both methods across all three deletion rates, achieving improvements of 0.100, 0.098, and 0.099 percentage over the strongest baseline at 40%, 60%, and 80% deletion, respectively. These results further demonstrate the effectiveness of MIRAGE under BFS-based edge deletion.

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

The paper studies graph domain adaptation when both source and target graphs have missing edges. MIRAGE constructs feature-based candidate edges and learns edge selection probabilities using reconstruction and intra-/cross-domain mutual-information objectives. Experiments on several benchmarks, including random and BFS-based edge deletion, report improvements over reconstruction and GDA baselines.

Critical strengths

The reported gains under severe edge deletion are substantial on several benchmarks. The additional BFS-based deletion experiments and transfer directions provide useful evidence that the results extend beyond a single random-deletion setting.

Critical weaknesses
  1. The significance and distinct contribution of the proposed setting are not sufficiently established. Missing edges are a reasonable robustness condition, but introducing this condition does not by itself establish a new adaptation problem requiring a dedicated framework. The paper needs a clearer distinction from existing structure-learning approaches to GDA, particularly the cited attribute-driven and homophily-agnostic structure-reconstruction work. Comparing primarily against conventional GDA and VGAE-based reconstruction leaves the contribution relative to these closer alternatives unresolved.

  2. The experiments do not isolate the claimed mechanism from a simpler feature-based explanation. Candidate edges are restricted to feature-similar node pairs, while features remain intact during edge deletion. A central alternative explanation is therefore that feature-based graph augmentation accounts for much of the improvement. The paper lacks a matched feature-kNN augmentation baseline and a systematic analysis of homophily/heterophily. The MI ablations show that additional objectives affect performance, but do not establish that they recover useful missing structure or specifically address a deletion-induced adaptation difficulty. The MAG results, where deletion can improve performance, further underline why this distinction matters.

Questions for authors
  1. What evidence distinguishes MIRAGE from simply adding feature-kNN edges before applying the same classifier or a strong GDA method? A comparison with matched candidate sets and edge budgets would materially affect my assessment.

  2. What supports the claim that missing edges introduce an additional adaptation difficulty, rather than ordinary changes in graph information and propagation? In particular, how should Figure 1 be interpreted given its comparison of source training loss with target loss, and how does this interpretation depend on homophily?

7 comments
AuthorsOpenReview ↗

Response to Reviewer dCWM

Part 1/4

We appreciate your constructive feedback. Below are our responses to your questions and concerns. Please feel free to reach out if further clarification is required.

1. Distinct Challenges of EIGDA and Comparison with Related Methods

We thank the reviewer for raising this concern. We would like to clarify the key challenge of EIGDA and why existing GDA methods may not effectively address it. Missing edges reduce the neighborhood information available for learning node representations, limiting their ability to capture local graph structures. More importantly, missing edges can affect source and target representations differently, and could increase the discrepancy between the two domains. Consequently, directly aligning representations learned from incomplete graphs may lead to inaccurate cross-domain alignment.

To further examine this challenge, we conducted additional experiments with two closely related methods, GAA [1] and RSGDA [2], under the same edge deletion settings. The results are reported in Table R1 and added to Appendix E (Table 7, in lines 976-978 on page 19, highlighted in blue) of the revised manuscript.

MethodB1→B2 60%B1→B2 80%A→D 60%A→D 80%B→U 60%B→U 80%
GAA0.3420.3230.531‾\underline{0.531}0.5120.348‾\underline{0.348}0.327‾\underline{0.327}
RSGDA0.360‾\underline{0.360}0.346‾\underline{0.346}0.5300.515‾\underline{0.515}0.3320.313
MIRAGE0.5100.4940.6180.5960.5480.539

Table R1. Accuracy comparison with GAA and RSGDA. The best results are in bold, and the second-best are underlined‾\underline{underlined}.

MIRAGE consistently outperforms both methods across all six settings, with average improvements of +0.149, +0.084 and +0.206 over the second-best on B1→B2, A→D, and B→U, respectively. These results highlight the limitations of their approaches under edge incompleteness. GAA jointly models attribute and topology representations for domain alignment, but its topology representations are affected by missing neighborhood information, which can limit the effectiveness of alignment. RSGDA constructs homophilic and heterophilic graph variants to model different structural patterns. However, missing edges also affect the structural information used to construct these variants, while its reconstruction focuses on modeling homophily and heterophily rather than addressing missing structural information. In contrast, MIRAGE learns which connections to directly add to the observed graphs, considering both the preservation of local neighborhood information and the effect of the selected connections on domain discrepancy.

AuthorsOpenReview ↗

Response to Reviewer dCWM

Part 2/4

2. Comparison with Feature-kkNN Augmentation Baselines

We would like to clarify that feature-kkNN is used only to construct candidate edge sets in MIRAGE, rather than to directly determine the reconstructed graph. Unlike simple feature-kkNN augmentation, which selects edges based solely on feature similarity, MIRAGE learns selection probabilities for candidate edges under mutual information objectives, which effectively mitigates domain discrepancy with the guidance of cross-domain MI.

To examine whether the improvements of MIRAGE can be explained by feature-based graph construction alone, we introduce two additional feature-kkNN baselines below. We have also added a description of these baselines in the revised manuscript (lines 307–310, page 6, highlighted in blue) and their details in lines 908-920 on page 17-18.

  • kkNN-GCN[3] constructs a sparse kk-nearest neighbor graph from the node feature matrix. Edges are then selected according to feature similarity, with the total number of edges matched to MIRAGE's final reconstructed graph for each domain. A GCN is trained on the constructed source graph and directly applied to the target graph for inference.
  • kkNN-Augmentation+GCN (kkNN-Aug+GCN for short) augments the observed graph with additional edges based on node feature similarity before GCN classification, following prior graph learning studies [3–10]. It uses the same candidate edge sets as MIRAGE and matches the number of added edges separately in the source and target domains.

We have We have added these two baselines to the main experimental tables (in lines 330-332, 348-350, and 367-368 on page 7), with the new results highlighted in blue. MIRAGE consistently outperforms both baselines across all evaluated datasets and deletion rates.

We further evaluate kkNN-Aug+GRADE, kkNN-Aug+StruRW and kkNN-Aug+HGDA and in Table R2 (also added to Appendix E, Table 7, in lines 978-980 on page 19, highlighted in blue) to examine whether feature-kkNN augmentation followed by conventional domain adaptation can achieve comparable performance.

MethodB1→B2 60%B1→B2 80%A→D 60%A→D 80%B→U 60%B→U 80%
kkNN-Aug+GRADE0.3290.3050.5310.4900.3630.352
kkNN-Aug+StruRW0.3710.348‾\underline{0.348}0.5140.4870.3250.307
kkNN-Aug+HGDA0.375‾\underline{0.375}0.3430.554‾\underline{0.554}0.523‾\underline{0.523}0.408‾\underline{0.408}0.396‾\underline{0.396}
MIRAGE0.5100.4940.6180.5960.5480.539

Table R2. Comparison with kkNN-Aug combined with three GDA methods. The best results are in bold, and the second-best are underlined‾\underline{underlined}.

As shown in Table R2, MIRAGE outperforms the second-best by an average of +0.117 in Accuracy across the six settings, with the largest average improvement on B1$\to$B2 (+0.142) and the smallest on A$\to$D (+0.069). These results demonstrate that MIRAGE's improvements cannot be explained by simply adding edges between nodes with similar features, even when combined with conventional domain adaptation.

AuthorsOpenReview ↗

Response to Reviewer dCWM

Part 3/4

3. Additional Adaptation Difficulty

We thank the reviewer for raising concerns about the additional adaptation difficulty caused by missing edges. In this response, we provide further clarification of this challenge and also address the reviewer's questions regarding the lack of a systematic homophily analysis.

We would like to further clarify the challenge of EIGDA from the perspective of graph domain adaptation. Beyond the reduced neighborhood information discussed in the paper, missing edges can cause source and target representations to deviate differently from those learned on complete graphs. As a result, directly aligning representations learned from incomplete graphs may establish misleading cross-domain alignment, which limits the effectiveness of adaptation. This additional challenge is reflected in both the source-target performance discrepancy and the structural discrepancy between the two domains, which we examine through the source-target loss gap (in lines 481-522 on page 9-10) and the source-target homophily gap (in lines 523-529 on page 10) in our revised manuscript, respectively.

Source-target loss gap. As shown in Fig. 6 (page 9, highlighted in blue), the source-target classification loss gap of GRADE increases with the edge deletion rate across all three transfer tasks, while MIRAGE consistently achieves a smaller gap. At 80% edge deletion, MIRAGE reduces the loss gap by 12.0%, 18.4%, and 30.4% on Blog1→Blog2, ACMv9→DBLPv7, and Brazil→USA, respectively, with an average reduction of 20.2%. This shows that MIRAGE helps mitigate the increased source-target classification performance discrepancy under edge incompleteness.

Source-target Homophily gap. To provide a systematic homophily analysis, we examine how the source-target homophily gap changes with edge deletion rates and compare the effects of MIRAGE reconstruction and feature-kkNN augmentation. As shown in Fig. 7 (page 10, highlighted in blue), the homophily gap of the observed graphs increases with the edge deletion rate across all three tasks, while MIRAGE consistently achieves a smaller gap than both the observed graphs and feature-kkNN augmentation. At 80% edge deletion, MIRAGE reduces the homophily gap by 52.3%, 42.4%, and 30.0% on the three tasks, respectively, compared with the observed graphs, with an average reduction of 41.6%. MIRAGE also achieves an average reduction of 36.1% compared with feature-kkNN augmentation.

These results show that missing edges can amplify both the classification performance discrepancy and the homophily discrepancy between source and target domains, while MIRAGE effectively mitigates both discrepancies through graph reconstruction.

4. Interpretation of Fig.1

SAGA [11] uses the source-target classification loss gap as an indicator of alignment effectiveness. Following this prior work, we use the loss gap to measure the performance discrepancy between source and target domains. In Fig. 1, the loss gap at 0% edge deletion reflects the original source-target performance discrepancy without missing edges. As the edge deletion rate increases, the loss gap of GRADE rises consistently across all three transfer tasks. Compared with the 0% setting, this increase indicates that missing edges introduce an additional source-target discrepancy beyond the original domain gap, highlighting the additional challenge they pose for domain adaptation.

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

This paper studies Edge-Incomplete Graph Domain Adaptation (EIGDA), where both source and target graphs have missing edges. The proposed method, MIRAGE, uses intra-domain and cross-domain mutual information to guide edge reconstruction, preserve local information, and support domain adaptation. Experiments on several datasets show improved target node classification under random and correlated edge deletion.

Critical strengths
  1. The paper studies a useful problem: domain adaptation when graph structures are incomplete. The motivation is clear, and the overall framework is easy to follow.

  2. The experiments include different deletion rates, BFS-based deletion, reconstruction followed by GDA, and component ablations. These results support the benefit of combining graph reconstruction with domain adaptation.

Critical weaknesses
  1. Figs. 2 and 5 show clearer class clusters, but the plots alone do not establish statistically significant improvement in source-target alignment. The t-SNE settings and stability across runs are unclear. Appendix I reports classification variability, which does not directly verify the visualization claims.

  2. The introduction mentions network crawling and omitted citations, and Section 5.5 evaluates correlated missing edges. However, a concrete application connecting these examples to the full EIGDA setting is still needed. The main experiments use the same deletion rate in both domains, leaving one-sided or unequal missingness less explored.

  3. The distinction from conventional graph completion needs a clearer explanation. Table 5 and the ablations support MIRAGE, but they do not fully isolate its reconstruction mechanism from the choice of adaptation objective. It is also unclear whether the method mainly recovers missing edges or adds new edges useful for classification.

Questions for authors
  1. Were source and target embeddings projected together in Figs. 2 and 5, with the same sampling and t-SNE settings across methods? Please report these settings and check stability across seeds. Can the authors also measure within-class source-target distance and class separation in the original embedding space, with uncertainty estimates over repeated runs?

  2. Could the authors describe one real-world example, specifying the source and target domains, prediction task, and why edges are missing while node features remain available? How does MIRAGE perform when only one domain has missing edges, or when the two domains have different deletion rates or missing-edge patterns?

  3. Can the authors compare MIRAGE with conventional reconstruction plus a standard alignment loss, using the same encoder, candidate edges, and added-edge budget? In Eq. (2), unobserved candidate edges are treated as negatives. How does the method overcome this penalty to recover missing edges? Reporting whether added edges match the deleted edges would help explain what the reconstruction achieves.

7 comments
AuthorsOpenReview ↗

Response to Reviewer cEkp

Part 1/2

We appreciate your constructive feedback! Below, we address the raised questions and concerns. Please feel free to reach out if further clarifications are needed.

1. Embedding Visualization and Quantitative Analysis (W1, Q1)

Were source and target embeddings projected together in Figs. 2 and 5, with the same sampling and t-SNE settings across methods? Please report these settings and check stability across seeds. Can the authors also measure within-class source-target distance and class separation in the original embedding space, with uncertainty estimates over repeated runs?

Thank you for the suggestion. We clarify that source and target embeddings are jointly projected into the same two-dimensional space using t-SNE. We use the same sampling procedure and t-SNE parameters across methods (n_components=2, random_state=33, perplexity=5, early_exaggeration=24). We also examine MIRAGE's visualizations across three random seeds in Fig. 10, in lines 1143-1157 on page 22 (highlighted in blue). As shown in Fig. 10, MIRAGE produces similar class clustering patterns across different runs.

To provide quantitative evidence beyond t-SNE visualizations, we further measure within-class source-target distance (lower is better) and class separation (higher is better) in the original embedding space. We report the mean and standard deviation over three runs in Table 10, in lines 1134-1142 on page 22 (highlighted in blue). MIRAGE achieves the lowest alignment distance (0.0064) and the highest class separation (0.6206), outperforming all compared methods on both metrics. These results further support the observations in Figs. 2 and 5.

2. Real-World Motivation and Different Edge Deletion Settings (W2, Q2)

Could the authors describe one real-world example, specifying the source and target domains, prediction task, and why edges are missing while node features remain available? How does MIRAGE perform when only one domain has missing edges, or when the two domains have different deletion rates or missing-edge patterns?

A practical example is research paper classification across academic citation networks, corresponding to our Citation dataset. The source domain consists of papers from ACM with known research categories, while the target domain consists of papers from DBLP whose categories need to be predicted. Citation links may be missing due to incomplete database coverage, citation extraction errors, or delayed updates. However, paper metadata and textual content, such as titles and abstracts, may still be available to construct node features.

We have evaluate MIRAGE under different edge deletion settings, including one-sided and unequal edge incompleteness, as well as different missing patterns. The results have been added to Tables 8-9, in lines 1080-1100 on page 21, highlighted in blue. MIRAGE still outperforms these baselines under all six edge deletion settings, achieving performance gains of at least +0.144 and +0.063 on Citation and Airport, respectively. This shows its effectiveness across different edge deletion settings.

AuthorsOpenReview ↗

Response to Reviewer cEkp

Part 2/2

3. Reconstruction Mechanism and Its Role in Domain Adaptation (W3, Q3)

  • Can the authors compare MIRAGE with conventional reconstruction plus a standard alignment loss, using the same encoder, candidate edges, and added-edge budget?

    To examine whether conventional graph reconstruction followed by domain alignment can achieve comparable performance, we introduce kkNN-Aug+GRADE, which adds edges based on node feature similarity before applying GRADE for domain adaptation. Details are provided in Appendix E. Both methods use a GCN encoder, and we match the candidate edge sets and the numbers of added edges in the source and target domains to those of MIRAGE. The results are reported in Table R1 (also included in lines 978-979 on page 19, highlighted in blue). MIRAGE outperforms kkNN-Aug+GRADE by an average of 0.185, 0.096, and 0.018 on Blog1→Blog2, ACMv9→DBLPv7, and Brazil→USA, respectively, across the two deletion rates. These results show that feature-based graph augmentation followed by conventional domain alignment does not achieve the same performance as MIRAGE's MI-guided reconstruction.

    MethodB1→B2 60%B1→B2 80%A→D 60%A→D 80%B→U 60%B→U 80%
    MIRAGE0.5100.4940.6180.5960.5480.539
    kkNN-Aug+GRADE0.3290.3050.5310.4900.3630.352

    Table R1. Comparison with kkNN-Aug+GRADE.

  • In Eq. (2), unobserved candidate edges are treated as negatives. How does the method overcome this penalty to recover missing edges? Reporting whether added edges match the deleted edges would help explain what the reconstruction achieves.

    Eq. (2) is not optimized independently. As shown in Eq. (5), MIRAGE jointly optimizes adjacency reconstruction and mutual information objectives. In particular, structure-level cross-domain MI directly depends on candidate-edge probabilities and provides additional gradients that can counteract the BCE penalty, allowing unobserved candidate edges to be selected.

    Importantly, MIRAGE aims to reconstruct graph structures that mitigates domain shift, rather than exactly recover the original topology. To further examine this distinction, we evaluate edge recovery F1, which measures the overlap between added edges and ground-truth deleted edges through precision and recall, on ACMv9→DBLPv7.

    VariantRecovery F1 60%Target Acc. 60%Recovery F1 80%Target Acc. 80%
    MIRAGE0.3330.6180.3060.596
    MIRAGE w/o cross-domain MI0.4600.5520.4010.535

    Table R2. Comparison of edge recovery F1 and target classification accuracy between MIRAGE and its variant without cross-domain MI under 60% and 80% edge deletion rates on ACMv9→DBLPv7.

    Removing cross-domain MI improves edge recovery F1 but reduces target accuracy by +0.066 and +0.061 at 60% and 80% deletion rates, respectively. These results show that higher overlap with the original edges does not necessarily lead to better domain adaptation, supporting the importance of transfer-aware reconstruction.

Reviewer_cEkpOpenReview ↗

Thanks

Thank you for the response and the additional experiments. My concerns about the visualization and the deletion settings are addressed, and the matched feature-kNN baselines show that the gain is not explained by feature-based augmentation alone.

However, I still have a remaining concern about the reconstruction mechanism. kNN-Aug+GRADE differs from MIRAGE in both the added edges and the adaptation objective, so could the authors report GRADE trained on MIRAGE’s reconstructed graphs? Second, Table R2 suggests that recovering deleted edges does not explain the performance gain, but it remains unclear what proportion of the added edges connect nodes of the same class.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary Comments

Dear AC/SAC/PC,

Thank you for your time and careful consideration of our submission. We also sincerely thank all reviewers for their thoughtful comments and constructive feedback throughout the rebuttal. As the discussion period comes to a close, we summarize the paper's contributions, major revisions, and discussion outcomes to assist your final assessment.

1. Contributions and Strengths

Our work introduces Edge-Incomplete Graph Domain Adaptation (EIGDA), where missing edges reduce the neighborhood information available for representation learning and can increase the discrepancy between source and target domains. To address this challenge, we propose MIRAGE, which uses intra-domain and cross-domain mutual information (MI) to guide graph reconstruction and mitigates domain shift.

We appreciate that Reviewer cEkp recognized EIGDA as a useful problem, with clear motivation and an easy-to-follow framework. Reviewer mWwq also recognized the importance of considering domain alignment during graph reconstruction, noting that the comparison with VGAE followed by GDA supports this motivation. Both reviewers acknowledged the extensive experimental evaluation, including multiple deletion rates, BFS-based deletion, reconstruction comparisons, and component ablations.

2. Summary of Revisions

  • Stronger baseline comparisons. We added GAA, RSGDA, feature-kkNN augmentation baselines combined with GCN and GDA methods, and GRADE trained on MIRAGE's reconstructed graphs.
  • Reconstruction mechanism. We analyzed added-edge homophily, edge recovery F1, reconstructed edge overlap across MI ablations, and the effect of MI components on source-target loss gaps.
  • Additional adaptation difficulty. We added classification loss gap and homophily gap analyses to show how missing edges increase source-target discrepancies and how MIRAGE mitigates them.
  • Additional experiments and analyses. We added experiments on different edge deletion settings, quantitative embedding analysis, fixed-threshold comparisons, further BFS-based deletion results, and an analysis of removed heterophilic edges on MAG.

The revisions and new experiments are highlighted in blue in the revised manuscript.

3. Summary of Discussion

  • Reviewer cEkp (3 → 4) acknowledged that our additional visualization, edge deletion analyses, and the matched feature-kkNN comparisons addressed key concerns. Following our additional clarifications and experiments on the reconstruction mechanism, the reviewer raised the score from 3 to 4.

  • Reviewer mWwq (3 → 4) explicitly acknowledged that the matched-edge comparisons with StruRW and HGDA and the added-edge analysis “address my main concerns”, and raised the score to 4 to support acceptance.

  • Regarding Reviewer dCWM (rating 1), the reviewer questioned whether EIGDA introduces a distinct adaptation challenge, whether MIRAGE's improvements can be explained by simple feature-based edge augmentation, and whether MI-guided reconstruction effectively addresses the difficulty caused by missing edges. In response, we first showed that missing edges amplify both the source-target loss gap and homophily gap, thereby introducing a distinct challenge for domain adaptation beyond changes in graph structure and message passing. At 80% edge deletion, MIRAGE reduces these gaps by an average of 20.2% and 41.6%, respectively, demonstrating its effectiveness in mitigating these discrepancies. We also added comparisons with GAA, RSGDA, and feature-kNN augmentation combined with conventional GDA methods, showing that MIRAGE consistently outperforms these alternatives. Furthermore, we conducted additional MI ablation analyses of reconstructed edge overlap and source-target loss gaps, demonstrating that MI components play a key role in guiding edge selection and reducing the source-target performance discrepancy under edge incompleteness.

Reviewer dCWM provided additional comments less than 30 minutes before the rebuttal deadline, leaving us insufficient time to fully address these new questions and engage in further discussion. We respectfully ask the Area Chair to consider the substantial experiments, analyses, and clarifications we have provided throughout the rebuttal when assessing the reviewer's concerns.

We sincerely thank the AC/SAC/PC and all reviewers again for their time, effort, and consideration. We hope this summary and the additional evidence will be helpful in the final assessment.

Best regards,

The Authors of Submission 18949

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.