acceptodds
Under review as a conference paper at ICLR 2027

Barlow Twins-Guided Adaptive Alignment for Multi-View Graph Domain Adaptation

Abstract

Multi-view graph domain adaptation (MGDA) transfers knowledge from a labeled source graph to an unlabeled target graph when both are observed through multiple views. Because domain shift differs across views, aligning the two graphs with a single view-agnostic objective is insufficient, yet existing methods weight all view pairs uniformly and overlook how each source view relates to its target counterpart. We characterize this relationship for every source–target view pair with two quantities: consistency and redundancy. Our empirical results reveal that domain discrepancy decreases with consistency and increases with redundancy. This suggests these two quantities can serve as both training signals and indicators of how strongly to align each pair. We therefore propose Barlow Twins-Guided Adaptive Alignment (BTGA), which applies the invariance and redundancy-reduction terms of Barlow Twins to each source–target view pair and uses the resulting consistency and redundancy scores to weight the pair's alignment. Experimental results on a variety of benchmark datasets verify the effectiveness of our method.

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

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.25 over 4 reviews, above 96% of the others.
Reviewer_4QrYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Barlow Twins-Guided Adaptive Alignment (BTGA) for unsupervised multi-view graph domain adaptation. It combines feature-aware graph filtering, a shared encoder, intra-domain Barlow Twins regularization, and adaptive cross-domain alignment. Shared PCA summaries enable correlation estimation between domains with unpaired nodes. Consistency and redundancy scores determine weights for aligning source-target view pairs through multi-kernel MMD. Training additionally uses source supervision and target entropy minimization. Experiments on twelve ACM and MAG transfer tasks report improvements over existing baselines. Ablation studies and efficiency comparisons provide supporting evidence, although several methodological and experimental inconsistencies require clarification.

Critical strengths
  1. The paper addresses a relevant limitation of multi-view graph adaptation: domain shifts vary across relation pairs, motivating differentiated alignment rather than uniform treatment.
  2. Reported results consistently outperform the evaluated baselines across twelve transfer tasks, with particularly substantial improvements on ACM and gains in both accuracy and macro-F1.
  3. The evaluation includes component ablations, parameter sensitivity analyses, and computational comparisons, providing useful evidence about the framework’s behavior beyond its main classification results.
Critical weaknesses
  1. The ablations omit uniform weighting with otherwise identical components. Removing alignment entirely cannot isolate the contribution of adaptive weighting, leaving the central methodological advantage over ordinary multi-view MMD alignment insufficiently demonstrated.
  2. Section 4.3 lists several hyperparameter choices but does not explain how they or the final checkpoint are selected without target labels.
  3. Figure 1 shows an association between the proposed scores and MMD on one transfer task. It does not establish why prioritizing already similar view pairs should improve target classification or when this strategy might fail.
  4. The paper reports five runs but provides no standard deviations or confidence intervals. This makes it difficult to judge the reliability of the smaller MAG gains and the differences between ablation variants.
Questions for authors

See the weaknesses.

5 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/2)

Thank you very much for your time and thoughtful feedback. We greatly appreciate your constructive and insightful suggestions.

W1. Comparison with uniform weighting.

Thank you for suggesting this comparison. We have added it to the ablation study in Section 4.5 of the main text (p. 8–9, lines 431–438, Figure 4) to assess the contribution of adaptive view-pair weighting.

We agree that w/o Alignment evaluates alignment as a whole and does not isolate adaptive weighting. We have therefore added ordinary multi-view MMD alignment as a uniform-weight control, assigning equal weights to all source–target view pairs while keeping other components unchanged. The table below reports some of the ACC results.

MethodACM2→ACM1US→CNJP→CNDE→CNFR→DEUS→RU
Ordinary multi-view MMD0.6410.5350.4930.4320.5360.460
BTGA0.6870.5580.5460.4610.5640.477

BTGA achieves higher ACC than uniform weighting on all twelve ACM and MAG tasks. These results support dynamically assigning alignment weights according to view-pair consistency and redundancy, enabling adaptive cross-domain alignment with better target classification. Full results with standard deviations are in Table E.4, Appendix E.4 (p. 25, lines 1327–1340).

W2. Hyperparameter and checkpoint selection.

Thank you for highlighting this issue. We have clarified the original selection protocol and added supplementary analyses without target-domain labels in Appendix B.3 (p. 17–20, lines 885–1051).

  • (i) A shared challenge in unsupervised domain adaptation. Hyperparameter and checkpoint selection is challenging without an independent labeled target-domain validation set. This challenge arises in both visual and graph domain adaptation.

  • (ii) Selection protocol for the main experiments. Following SAGA's emphasis on consistent training and evaluation [1], we compare BTGA and the baselines under the same data partitions and evaluation conditions. BTGA's hyperparameters are selected using target-domain test performance, and its reported checkpoint is the epoch with the highest target ACC within the predefined training budget. Target labels are used for hyperparameter selection, reporting-epoch selection, and evaluation, but not model parameter optimization.

  • (iii) Supplementary selection without target-domain labels. We have added Soft Neighborhood Density (SND) [2] and Importance Weighted Cross-Validation (IWCV) [3] analyses. SND maximizes a neighborhood-density score from unlabeled target predictions; IWCV minimizes an importance-weighted source-validation loss.

    We independently vary α\alpha and β\beta over 0.25,0.5,1,2,4\\{0.25,0.5,1,2,4\\} times their task-specific defaults, keeping other hyperparameters fixed. Candidates are scored and selected using fixed final checkpoints, without target-domain labels. The supplementary analysis therefore does not select the reporting epoch by target ACC. The table below reports some of the ACC results.

    CriterionTaskParameterMultiplierSelected ACCMain ACC
    SNDDE→FRβ\beta0.5×0.5570.560
    IWCVCN→USα\alpha1×0.5880.592

    The selected configurations achieve ACC within 0.3 and 0.4 percentage points of the corresponding main results. On the representative transfer tasks presented, both criteria identify configurations whose classification performance is close to that reported in the main experiments. These results support the effectiveness of SND and IWCV for selecting BTGA’s α\alpha and β\beta within the examined candidate ranges without using target-domain labels. Further selection analyses are provided in Figures B.2–B.3 (p. 19–20, lines 1014–1017 and 1043–1046).

References

[1] Fang et al. SAGA: Structural aggregation guided alignment with dynamic view and neighborhood order selection for multiview graph domain adaptation. ICLR 2026.

[2] Saito et al. Tune it the right way: Unsupervised validation of domain adaptation via soft neighborhood density. ICCV 2021.

[3] Sugiyama et al. Covariate shift adaptation by importance weighted cross validation. JMLR 2007.

AuthorsOpenReview ↗

Official Comment by Authors (2/2)

W3. Rationale for prioritizing view pairs with smaller domain discrepancies.

Thank you for highlighting this concern. We explain the rationale and empirical evidence below, including performance when view-pair matching deteriorates. The strategy comparison has been added to Section 4.5 of the main text (p. 8–9, lines 431–438, Figure 4).

  • (i) Rationale and empirical evidence. Our aim is to prioritize information useful for target classification. We believe that smaller-discrepancy view pairs offer a more reliable basis for transfer. BTGA therefore gives them greater relative weights while retaining all pairs in alignment.

    We fix the task and target view, retaining the full model and other losses. From the same state after five epochs without alignment, we compare no alignment with aligning either source view (coefficient 1). The table below reports some of the results.

    Task (target view)Aligned pairMK-MMD²Target ACCNo-alignment ACC
    ACM2→ACM1 (PAP)PAP→PAP0.53450.5280.434
    ACM2→ACM1 (PAP)PSP→PAP0.80000.4710.434
    DE→CN (PP)PP→PP0.38810.4690.423
    DE→CN (PP)PAP→PP0.39180.4380.423

    Both alignment settings improve upon no alignment, but the smaller-discrepancy pair yields higher ACC by 5.7 and 3.1 percentage points, respectively. Thus, larger discrepancy does not necessarily imply greater alignment benefit. Full results are in Table G.3, Appendix G.3 (p. 30, lines 1600–1610).

  • (ii) Comparison with large-discrepancy prioritization. In Section 4.5 of the main text (p. 8–9, lines 431–438, Figure 4), we compare BTGA with a control assigning higher weights to larger-discrepancy view pairs, keeping other components unchanged. The table below reports some of the ACC results.

    MethodACM2→ACM1US→CNJP→CNDE→CNFR→DEUS→RU
    Large-discrepancy prioritization0.5660.4330.4920.4280.5210.424
    BTGA0.6870.5580.5460.4610.5640.477

    BTGA achieves higher ACC on all twelve tasks, supporting its strategy of dynamically assigning greater relative weights to high-consistency, low-redundancy view pairs. Full results with standard deviations are in Table E.4, Appendix E.4 (p. 25, lines 1327–1340).

  • (iii) When the strategy may fail. Alignment weights express relative priorities among available pairs. When all pairs are poorly matched, reweighting alone cannot ensure reliable transfer. We progressively increase target-attribute Gaussian noise using σ∈0,2,4,6\sigma\in\\{0,2,4,6\\}, scaled by each feature's standard deviation, and retrain BTGA at each level (Appendix G.3, p. 31, lines 1622–1631). The table below reports some of the ACC results.

    Noise multiplier σ\sigmaDE→FRCN→JPJP→CNRU→US
    0 (clean)0.5450.5330.4940.387
    20.4610.4280.4330.338
    40.3740.3190.3460.275
    60.3350.2390.2940.227

    ACC declines as noise increases on these tasks (Appendix G.3, Figure G.5(b), p. 31, lines 1652–1655). At σ=4\sigma=4, all four view pairs in each task show lower consistency and higher redundancy than their clean counterparts. These results illustrate a limitation: relative weighting does not prevent classification degradation when all available view pairs deteriorate.

W4. Statistical reliability of experimental results.

Thank you for highlighting the need to report variability. We have added mean ± standard deviation across five runs for BTGA's main results, component ablations, and filtering/alignment comparisons in Appendix E.4, Tables E.1–E.4 (p. 24–25, lines 1273–1340). The table below reports some of the results.

Method / metricACM1→ACM2CN→USDE→FRRU→US
BTGA / ACC0.666 ± 0.0190.592 ± 0.0040.560 ± 0.0040.430 ± 0.013
BTGA / Macro-F10.664 ± 0.0130.410 ± 0.0050.494 ± 0.0050.418 ± 0.017
w/o Alignment / ACC0.361 ± 0.0090.554 ± 0.0080.544 ± 0.0050.384 ± 0.009
Consistency Only / ACC0.383 ± 0.0200.563 ± 0.0280.534 ± 0.0130.384 ± 0.016
Redundancy Only / ACC0.384 ± 0.0100.564 ± 0.0080.523 ± 0.0270.383 ± 0.005

These additions make run-to-run variability explicit for the MAG results and ablation variants. We assess the reported mean differences together with these SDs, providing a more transparent basis for evaluating their reliability.

AuthorsOpenReview ↗

Looking forward to hearing from you

Dear Reviewer 4QrY,

We deeply appreciate your time and effort in reviewing our submission and providing valuable feedback. Your insights are crucial to our work.

Following your suggestions, we have added a uniform-weighting comparison, clarified hyperparameter and checkpoint selection, explained the alignment priority and its limitations, and reported means and standard deviations across five runs. These additional analyses and experiments have greatly improved the quality of our paper.

As the discussion deadline approaches, we would be grateful for your feedback and will do our best to address any remaining concerns.

Best regards,
The Authors of Submission 31342

Reviewer_BK59OpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

BTGA addresses multi-view graph domain adaptation by characterizing each source–target view pair with consistency and redundancy. It uses Barlow Twins to reduce intra-domain cross-view redundancy, derives pair-specific alignment weights, and applies weighted MK-MMD alignment. Experiments on ACM and MAG across twelve transfer tasks show consistent gains over GDA/MGDA baselines, with ablations supporting graph filtering, intra-domain redundancy reduction, and adaptive weighting.

Critical strengths

First, the central idea that weighting source–target view pairs by consistency and redundancy instead of aligning all views equally is interpretable and well-motivated. This is consequential because it directly addresses a real limitation of MGDA: view-level domain discrepancy varies across view pairs. The empirical gains are substantial and consistent, and improved average ACC/Macro-F1 on all ten MAG tasks.

Second, the method is computationally practical: it achieves competitive runtime with the lowest peak GPU memory among compared methods, while ablations isolate the contributions of graph filtering, intra-domain redundancy reduction, and adaptive alignment.

Critical weaknesses

The motivation for using Barlow Twins is insufficiently explained for cross-domain view-pair weighting. Barlow Twins was designed for two augmented views of the same sample; applying diagonal/off-diagonal correlations to source–target view pairs from different domains lacks justification for why these statistics indicate transferability.

Questions for authors

Please clarify the construction of the positive/negative sigma summaries HD,v∈R2r×dH^{D,v}\in\mathbb{R}^{2r\times d}: what equations define the 2r2r rows, how are the rr shared principal directions obtained, how is rr selected, and what temperature τ\tau is used? Are these summaries recomputed each epoch, and how sensitive are the alignment weights to this refresh rate?

Also, the manuscript does not appear to mention source-code release; given the complex weighting mechanism and task-specific hyperparameters, please confirm whether code will be made available and provide enough implementation detail for independent reproduction.

Finally, please justify why Barlow Twins statistics are appropriate for cross-domain view pairs.

4 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/2)

We greatly appreciate your thorough review and constructive feedback. We are pleased to address the points you raised.

W1 & Q3: Barlow Twins for cross-domain view-pair weighting.

Thank you for highlighting this concern. We have added theoretical motivation and experimental evidence in Section 4.7 of the main text (p. 9–10, lines 463–495) and Section 4.5 (p. 8–9, lines 398–435) to explain why BT statistics are used to determine cross-domain view-pair weights.

  • (i) Theoretical motivation. GAA [1] shows that the target-risk bound depends on node-attribute discrepancies ∥XiS−XjT∥22\|X_i^S-X_j^T\|_2^2 and structure-aggregated feature discrepancies ∥(ASXS)i−(ATXT)j∥22\|(A^SX^S)_i-(A^TX^T)_j\|_2^2, relating to attribute consistency and structure-related consistency, respectively. HGDA [2] likewise incorporates attribute and structural-signal discrepancies into its target-risk bound. Under the respective assumptions and with other terms fixed, reducing these discrepancies can tighten the risk bounds, motivating transfer through better-matched attribute and structural information.

    BTGA's relation-view representations combine node attributes and relational structure in a shared feature space. CC measures agreement between corresponding feature dimensions, providing a basis for comparing cross-domain matching.

    Corresponding-dimension agreement can coexist with repeated encoding of shared signals across dimensions. Barlow Twins [3] combines corresponding-dimension agreement with cross-dimensional redundancy reduction. BTGA uses RR to measure squared cross-domain correlations between different dimensions, capturing redundancy not reflected by CC. We believe that maintaining consistency while reducing redundant encoding helps identify view pairs more suitable for transfer and target classification. BTGA therefore jointly uses CC and RR to dynamically assign greater relative weights to high-consistency, low-redundancy pairs.

  • (ii) Experimental evidence. We examine Consistency–MMD and Redundancy–MMD across all twelve ACM and MAG tasks. The table below reports some of the results as Pearson correlations over all observation points on each plotted four-pair mean curve (Figure 6, p. 9, lines 483–485).

    Relationship (Pearson r)ACM1→ACM2ACM2→ACM1US→CN
    Consistency–MMD−0.975−0.962−0.729
    Redundancy–MMD0.9930.9940.988

    Higher consistency generally accompanies smaller MMD, whereas higher redundancy generally accompanies larger MMD, supporting their use to evaluate view-pair domain discrepancy. Results for the other nine tasks are in Appendix G.2 (p. 29, lines 1539–1549; Figure G.4, p. 30, lines 1596–1599).

    We further compare joint C/R weighting with Consistency Only and Redundancy Only, keeping the other components unchanged. The table below reports some of the ACC results from the tasks in Figure 3 (p. 8, line 419).

    Weighting strategyACM1→ACM2ACM2→ACM1RU→USCN→DE
    Consistency Only0.3830.4240.3840.354
    Redundancy Only0.3840.4210.3830.352
    BTGA0.6660.6870.4300.412

    BTGA achieves higher ACC than both variants on all twelve tasks. These results support jointly using BT's consistency and redundancy statistics to determine cross-domain view-pair weights and improve target classification. Full results with standard deviations are in Table E.2, Appendix E.4 (p. 25, lines 1296–1312).

BTGA uses BT's correlation criteria for view-pair evaluation without assuming that source and target nodes are paired augmentations. The computation for unpaired domains is detailed in our response to Q1 and Appendix H.2 (p. 32, lines 1676–1695).

References

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

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

[3] Zbontar et al. Barlow Twins: Self-Supervised Learning via Redundancy Reduction. ICML 2021.

AuthorsOpenReview ↗

Official Comment by Authors (2/2)

Q1. Construction of the positive/negative sigma summaries and sensitivity to the refresh rate.

Thank you for requesting these implementation details. We have expanded the construction formulas, parameter settings, refresh procedure, and sensitivity analysis in Appendix H.2 (p. 32, lines 1676–1727, Figure H.1). We address the six points below.

1. Definition of the positive/negative summary rows. For domain D∈S,TD\in\\{S,T\\} with NDN_D nodes, concatenate its view representations as JD=[ZD,1,…,ZD,V]J^D=[Z^{D,1},\ldots,Z^{D,V}] and denote the column-centered result by XDX^D. Let uku_k be the shared direction defined in point 2. For k=1,…,rk=1,\ldots,r, we define

\sigma_{D,k} &=\sqrt{\max\left\\{ \frac{1}{N_D}\sum_{n=1}^{N_D}(X^D_{n,:}u_k)^2,\epsilon \right\\}},\\\\ H^D_{k,:}&=\sigma_{D,k}u_k^\top,\qquad H^D_{r+k,:}=-\sigma_{D,k}u_k^\top. \end{aligned}$$ Here, $\epsilon=10^{-5}$. Positive/negative denotes opposite signs along a shared direction, not class labels. We split $H^D\in\mathbb R^{2r\times Vd}$ into view-specific column blocks $H^{D,v}\in\mathbb R^{2r\times d}$. Corresponding rows across domains share the same direction and sign, without pairing original nodes. Domain means are not added back. Before computing correlations, each column is centered and divided by $\sqrt{\operatorname{Var}+\epsilon}$, using population variance over its $2r$ rows. **2. Shared principal directions.** We compute each domain's covariance and their equally weighted average: $$\Sigma_D=\frac{(X^D)^\top X^D}{N_D},\qquad \overline{\Sigma}=\frac{\Sigma_S+\Sigma_T}{2}.$$ The eigenvectors of $\overline{\Sigma}$, ordered by decreasing eigenvalue, define $u_1,\ldots,u_r$. Both domains use the same directions and ordering. Each eigenvector's largest-magnitude entry is made positive to fix its sign. **3. Choice of $r$.** Across ACM and MAG, we use $V=2$ views and representation dimension $d=128$ per view. We retain all joint directions, giving **$r=Vd=256$**. **4. Temperature.** We use **$\tau=0.1$** across ACM and MAG to convert view-pair scores into normalized alignment weights. **5. Whether and when summaries are refreshed.** Full-domain summaries are computed without target labels, initialized before training, and refreshed at the end of every epoch for use in the next epoch. The statistics are detached from the encoder, while the redundancy coefficient continues to be optimized. Weights are recalculated from the cached statistics and current coefficient. **Caching statistics therefore does not freeze the alignment weights.** **6. Sensitivity to the refresh rate.** We compare intervals $K=1,2,5,10$ across all twelve tasks using seed 0 (**Figure H.1**, p. 32, lines 1702–1727). Training weights are first averaged over batches within each epoch. Weight MAE then averages absolute differences from the $K=1$ reference over the four view pairs and all epochs. Longer intervals generally produce larger weight deviations (**Figure H.1(a)**, p. 32, lines 1722–1723). The ACM2→ACM1 example shows delayed early weight redistribution (**Figure H.1(b)**, p. 32, lines 1723–1724). Increasing $K$ from 1 to 10 changes peak target ACC from **0.6985 to 0.6965**, and Macro-F1 at the corresponding peak-ACC checkpoints from **0.6825 to 0.6790** (**Figure H.1(c)**, p. 32, lines 1724–1727). Classification changes are small in this illustrated case, while the default **$K=1$** keeps the domain-level statistics up to date. Task-specific training settings and the training algorithm are provided in **Appendices B and D** (Table B.1, p. 17, lines 864–879; Algorithm 1, p. 22, lines 1136–1176). > **Q2. Code availability and reproducibility.** Thank you for highlighting the importance of code availability and reproducibility. We will make our source code publicly available upon acceptance of the paper to help other researchers reproduce our results.

Thank you for the clarifications. Since the response has addressed my concerns, I keep my rating unchanged.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you so much for confirming that our response has addressed your concerns and for your continued support of our work. We sincerely appreciate your careful review and constructive suggestions, which have helped us improve the clarity and rigor of our paper. Thank you again for your time and thoughtful engagement throughout the discussion.

Reviewer_4siaOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes BTGA for multi-view graph domain adaptation. It combines feature-aware graph filtering, intra-domain Barlow Twins regularization, and adaptive cross-domain alignment. Consistency and redundancy scores determine the weights of source-target view pairs in an MK-MMD loss. Experiments on ACM and MAG report improvements over the compared baselines.

Critical strengths
  1. The paper addresses a relevant problem: different relation views can have different domain shifts. Using view-pair statistics to guide alignment is a reasonable idea.

  2. Experiments cover twelve transfer tasks on ACM and MAG. The component ablations, including consistency-only and redundancy-only weighting, help assess the proposed design.

Critical weaknesses
  1. The relationship between consistency, redundancy, and domain discrepancy needs a clearer explanation. In Figure 1(b), MMD rises again at high consistency, so the relationship is not monotonic. It is also unclear whether reducing redundancy always benefits transfer or whether useful information could be lost.

  2. The presentation does not clearly explain the different roles of these two quantities within and across domains. Eq. (5) includes both consistency and redundancy terms, although the section emphasizes redundancy reduction. Across domains, both quantities determine alignment weights, but they are not directly optimized as a cross-domain Barlow Twins loss. The motivation for this distinction needs clarification. ACM and MAG use different relation views and have different graph statistics. Although Appendix B reports parameter sensitivity, the paper does not clearly connect these dataset differences to measured consistency, redundancy, or the preferred parameter settings.

Questions for authors
  1. How do consistency and redundancy interact? Why does MMD increase again at high consistency in Figure 1(b)? Could the authors examine each quantity while controlling for the other, and clarify whether lower redundancy consistently improves target classification?

  2. Why are consistency and redundancy directly regularized within domains but used only to weight MK-MMD across domains? Could the authors explain this choice and compare it with directly applying both terms across domains?

  3. Could the authors report consistency and redundancy statistics for ACM and MAG and relate them to their graph properties, such as relation types and edge density? How do these differences affect the learned alignment weights and parameter choices?

  4. Are the representations in Eq. (4) centered before computing the correlations? Please clarify its relationship to Eq. (7), distinguish the two matrices explicitly, and use consistent loss notation.

4 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/4)

Thank you very much for your insightful review and valuable feedback on our paper. Below, we respond to each of your comments.

W1 & Q1: Consistency, redundancy, and domain discrepancy.

Thank you for pointing out the non-monotonic behavior in Figure 1(b) (p. 1, lines 49–53). We have extended the analysis to all twelve transfer tasks on ACM and MAG, examining how consistency and redundancy each relate to MMD. These results have been added to Section 4.7 of the main text (p. 9–10, lines 463–495, Figure 6).

  • (i) Joint roles of consistency and redundancy. CC measures corresponding-dimension agreement in source–target statistical summaries; RR averages squared cross-dimensional correlations. High CC can coexist with repeated encoding, while low RR does not ensure corresponding-dimension agreement. BTGA assigns larger relative weights to smaller scores sp,q=(1−Cp,q)+λinterRp,qs_{p,q}=(1-C_{p,q})+\lambda_{\mathrm{inter}}R_{p,q}. With other conditions fixed, higher CC or lower RR increases its weight. Across tasks, higher CC generally accompanies lower MMD, whereas higher RR generally accompanies higher MMD (Figure 6, p. 9, lines 483–485). See Appendix G.2 (p. 30, lines 1596–1599, Figure G.4).

  • (ii) Why MMD can rise again at high consistency. We acknowledge that the consistency–MMD relationship is not strictly monotonic. The blue curve for US→CN in Figure 6(c) (p. 9, lines 483–485) also shows an MMD rebound at high consistency. Consistency is based on correlations between corresponding feature dimensions, whereas MMD measures the discrepancy between source and target feature distributions. Thus, high consistency alone does not guarantee that the distributions are well aligned.

    We believe excessive emphasis on consistency risks overfitting to shared information and overlooking useful complementary information across views. Wang et al. [1] show that retaining only shared information can discard non-shared task-relevant information and impair downstream performance. Shen et al. [2] further identify view-unique task-relevant information in multiplex graphs. These studies support preserving complementary information alongside shared information.

  • (iii) Whether lower redundancy always benefits classification. Reducing redundancy alone need not improve target ACC and can risk losing useful information. RR measures feature correlations, not class separability or task relevance. We believe suppressing correlations carrying class information can weaken discriminability. Zhao et al. [3] provide a counterexample showing that domain-invariant representations and low source classification error do not guarantee good target classification. Chen et al. [4] show that adversarial alignment can improve transferability at the expense of feature discriminability. These studies emphasize preserving discriminative information during distribution alignment.

We compare Consistency Only, Redundancy Only, and full BTGA, retaining all other components, including source supervision. The table below reports some of the ACC results.

Weighting strategyACM1→ACM2ACM2→ACM1CN→USUS→CNDE→FRFR→DE
Consistency Only0.3830.4240.5630.5240.5340.513
Redundancy Only0.3840.4210.5640.5210.5230.514
BTGA0.6660.6870.5920.5580.5600.564

BTGA achieves higher ACC than both variants on all twelve tasks, supporting jointly using consistency and redundancy to guide adaptive view-pair alignment. See Section 4.5 of the main text (p. 8–9, lines 398–435, Figure 3); full results with SDs are in Table E.2, Appendix E.4 (p. 25, lines 1296–1312).

References

[1] Wang et al. Rethinking Minimal Sufficient Representation in Contrastive Learning. CVPR 2022.

[2] Shen et al. Beyond Redundancy: Information-aware Unsupervised Multiplex Graph Structure Learning. NeurIPS 2024.

[3] Zhao et al. On Learning Invariant Representations for Domain Adaptation. ICML 2019.

[4] Chen et al. Transferability vs. Discriminability: Batch Spectral Penalization for Adversarial Domain Adaptation. ICML 2019.

AuthorsOpenReview ↗

Official Comment by Authors (2/4)

W2 & Q2: Different roles of Barlow Twins within and across domains.

Thank you for highlighting this distinction. We have clarified the roles of consistency and redundancy within and across domains and added two replacement experiments in Section 4.7 of the main text (p. 10, lines 496–504).

  • (i) Why intra-domain BT is used for regularization. Barlow Twins jointly encourages cross-view agreement and reduces redundant feature correlations [5]. Graph Barlow Twins adopts this principle for graph representation learning [6], and BTGF applies it to multiple relation views [7]. Within each domain, relation views contain the same nodes, allowing direct regularization of their corresponding representations. The diagonal term encourages agreement between corresponding feature dimensions of the same nodes; the off-diagonal term reduces correlations between different dimensions. Thus, Eq. (5) (p. 4, lines 191–194) contains both consistency and redundancy terms: together they maintain cross-view agreement, reduce repeated encoding, and help avoid representation collapse. Further discussion is in Appendix H.1 (p. 31, lines 1660–1665).

  • (ii) Why cross-domain BT is used for weighting. Across domains, we need an alignment loss to match source and target representation distributions. The BT loss regularizes feature correlations, and minimizing it alone does not ensure distribution matching. We therefore use consistency and redundancy derived from BT to evaluate source–target view pairs and determine their relative weights in the alignment loss. We clarify this choice in Appendix H.1 (p. 31, lines 1666–1673) and compare it with directly using a cross-domain BT loss in the experiments below.

  • (iii) Validation through two replacement experiments. We compare BTGA with two variants: replacing weighted cross-domain MK-MMD with a Barlow Twins loss while retaining intra-domain BT; and replacing intra-domain BT with MK-MMD while retaining the original cross-domain weighting and alignment. The table below reports some of the target-domain results; each cell gives ACC / Macro-F1.

    MethodACM1→ACM2ACM2→ACM1CN→USUS→CNDE→FRFR→DE
    Cross-domain MK-MMD → BT loss0.608 / 0.5780.609 / 0.5840.528 / 0.3500.489 / 0.4310.501 / 0.4300.497 / 0.425
    Intra-domain BT → MK-MMD0.612 / 0.6060.648 / 0.6350.570 / 0.3880.535 / 0.4320.538 / 0.4550.540 / 0.470
    BTGA0.666 / 0.6640.687 / 0.6840.592 / 0.4100.558 / 0.4550.560 / 0.4940.564 / 0.493

    BTGA achieves higher ACC and Macro-F1 than both variants on all twelve tasks. The two replacements reduce task-averaged ACC by 6.75 and 3.12 percentage points, respectively. These results support using BT to regularize within-domain cross-view representations and C/R to weight cross-domain distribution alignment.

    The comparison is summarized in Section 4.7 of the main text (p. 10, lines 496–504). Settings and full results are in Appendix H.3 (p. 33, lines 1730–1736; Figure H.2, lines 1754–1755) and Appendix H.4 (p. 33, lines 1759–1768; Figure H.3, p. 34, lines 1799–1800).

References

[5] Zbontar et al. Barlow Twins: Self-Supervised Learning via Redundancy Reduction. ICML 2021.

[6] Bielak et al. Graph Barlow Twins: A Self-Supervised Representation Learning Framework for Graphs. Knowledge-Based Systems, 2022.

[7] Qian et al. Upper Bounding Barlow Twins: A Novel Filter for Multi-Relational Clustering. AAAI 2024.

AuthorsOpenReview ↗

Official Comment by Authors (3/4)

W2 & Q3: Graph properties, C/R, alignment weights, and parameter settings.

Thank you for asking how dataset characteristics relate to our mechanism. We have added graph properties and C/R–weight statistics for all twelve ACM and MAG tasks, together with an analysis of filter-parameter choices, in Appendix G.1 (p. 26–29, lines 1361–1537).

  • (i) How graph properties enter the view representations. ACM's PAP and PSP views connect papers through shared authors and subjects, respectively. Their densities are 0.2870%–0.3334% and 24.1346%–26.8399%. MAG's PAP and citation-based PP views are both sparse (0.0134%–0.0516%). Within each domain, relation views share node attributes but aggregate them through different neighborhoods. Filtering and encoding therefore produce relation-specific representations, whose cross-domain agreement and redundancy are assessed by C/R from standardized summaries. Input densities and raw attribute-norm distributions are shown in Figure G.1 (p. 26, lines 1389–1400).

    To connect structure and attributes more directly, Figure G.2 (p. 27, lines 1418–1424) compares source–target feature-homophily differences with initial weights in eight selected illustrative directions. In each selected direction, the two pairs with smaller differences receive higher initial weights than the two with larger differences. Feature homophily measures mean attribute cosine similarity over non-self edges. This complements edge density by describing the attribute patterns connected by each relation.

  • (ii) How measured C/R determines alignment weights. Table G.1 (p. 27–28, lines 1440–1494) reports initial C/R, initial weights, and final weights for all twelve tasks using seed 0. C/R and initial weights are measured before training; final weights come from the last epoch of the same run. Representative rows are:

    TaskView pairInitial CInitial RInitial weightFinal weight
    ACM1→ACM2PAP→PAP0.99650.005750.64360.7249
    ACM1→ACM2PSP→PSP0.91030.007120.26820.1440
    CN→USPAP→PAP0.49160.003860.34740.3183
    CN→USPAP→PP0.37010.002550.10440.1819

    On ACM1→ACM2, the sparser authorship pair has higher C and lower R than the denser subject-based pair; both favor its larger initial weight. On CN→US, PAP→PAP has both higher C and higher R than PAP→PP, yet receives more weight. Here, the consistency gain outweighs the redundancy penalty in the joint score:

    \omega_{p,q}\propto\exp(-s_{p,q}/\tau).$$ These cases illustrate reinforcing and counterbalancing contributions of C/R in the initial allocation; the final-weight column separately records the learned allocation. BTGA evaluates matching through their joint score rather than assigning weights from density or redundancy alone. The coefficient $\lambda_{\mathrm{inter}}$ is learned, while $\tau$ controls weight concentration.
  • (iii) Connection to parameter choices. The filter order kk controls structural propagation, while γ\gamma balances attribute information and topology-based filtering. The adopted settings use k=6k=6 on ACM and k=1k=1–44 on MAG (Table B.1, p. 17, lines 864–879). Table G.2 (p. 28, lines 1503–1511) reports C/R under representative adopted configurations.

    We further compare k=1,2,4,6k=1,2,4,6 on CN→US with seed 0, fixing γ=100\gamma=100, encoder initialization, and other settings. C/R is averaged over the four view pairs before training; ACC is measured after 50 epochs.

    Filter order kInitial CInitial RFinal ACC
    10.47840.004290.5599
    2 (adopted)0.43770.003720.5901
    40.40200.003260.5821
    60.39340.003140.5820

    The highest C occurs at k=1k=1 and the lowest R at k=6k=6, while k=2k=2 achieves the highest ACC. From k=1k=1 to 22, both C and R decrease, yet ACC improves by 3.02 percentage points; from k=2k=2 to 66, R decreases further, but ACC drops by 0.81 percentage points. The adopted k=2k=2 therefore illustrates a balance between retaining consistency and reducing redundancy. This comparison is reported in Appendix G.1 (p. 28–29, lines 1496–1518; Figure G.3, p. 29, lines 1532–1535). The main-experiment selection protocol is in Appendix B.3 (p. 17, lines 886–895).

AuthorsOpenReview ↗

Official Comment by Authors (4/4)

Q4: Centering, correlation matrices, and loss notation.

Thank you for requesting these clarifications. Yes, the representations used in Eq. (4) are centered and standardized across nodes before computing correlations. We have revised the matrix notation and standardization description in Sections 3.3–3.4 of the main text (p. 4–5, lines 179–221), explicitly distinguishing the intra-domain matrix ΨD,pq\Psi^{D,pq} from the cross-domain matrix Mp,q\mathbf M^{p,q}.

  • (i) Centering and standardization. For an input matrix Y∈Rm×dY\in\mathbb R^{m\times d}, each feature column is standardized as Y^ℓ,i=Yℓ,i−μivi+ϵ,\widehat Y_{\ell,i}=\frac{Y_{\ell,i}-\mu_i}{\sqrt{v_i+\epsilon}}, where μi\mu_i and viv_i are its mean and population variance across rows. Intra-domain BT uses batch normalization without learnable affine parameters. Cross-domain statistical summaries are also centered and standardized column-wise, using vi+ϵ\sqrt{v_i+\epsilon} for numerical stability. Intra-domain preprocessing is explained in Appendix H.1 (p. 31, lines 1660–1665), and summary standardization in Appendix H.2 (p. 32, lines 1693–1695).

  • (ii) Relationship between Eqs. (4) and (7). Eq. (4) (p. 4, lines 183–190) defines the intra-domain matrix; Eq. (7) (p. 5, lines 218–221) defines the cross-domain matrix. With standardized inputs, the implementation computes

    \mathbf M^{p,q}=\frac{1}{2r}(\widetilde H^{S,p})^\top\widetilde H^{T,q}.$$ Here, $B_D$ is the number of nodes in the intra-domain batch, and $2r$ is the number of statistical-summary rows. Both matrices are $d\times d$ and calculate feature cross-correlations, but their inputs and row correspondences differ. | Property | Intra-domain matrix $\Psi^{D,pq}$ | Cross-domain matrix $\mathbf M^{p,q}$ | |---|---|---| | Inputs | Node representations from two views of the same domain | Source and target statistical summaries | | Row correspondence | The same node across relation views | The same shared principal direction and sign | | Role | Construct the intra-domain BT loss | Estimate C/R for MK-MMD weighting | The summaries use shared principal directions and matching positive/negative signs. Cross-domain correlation therefore uses statistical row correspondences, without requiring one-to-one source–target node matches or equal numbers of nodes. Their construction is detailed in **Appendix H.2** (p. 32, lines 1676–1695).
  • (iii) Consistent loss notation. ΨD,pq\Psi^{D,pq} and Mp,q\mathbf M^{p,q} denote correlation matrices. LBT\mathcal L_{\mathrm{BT}} denotes one intra-domain view-pair loss; Lintra\mathcal L_{\mathrm{intra}} sums the within-domain view-pair averages for the source and target domains; and Lalign\mathcal L_{\mathrm{align}} denotes C/R-weighted cross-domain MK-MMD. The total objective is L=LS+Lintra+αLalign+βLT.\mathcal L=\mathcal L_S+\mathcal L_{\mathrm{intra}}+\alpha\mathcal L_{\mathrm{align}}+\beta\mathcal L_T. Here, LS\mathcal L_S is the source classification loss and LT\mathcal L_T the target entropy loss. These definitions follow Eqs. (5)–(6) (p. 4, lines 191–206), Eq. (12) (p. 5, lines 258–260), and Eq. (15) (p. 6, lines 277–283). The different roles of the two matrices are discussed in Appendix H.1 (p. 31, lines 1660–1673).

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

This paper addresses the issue in multi-view graph domain adaptation (MGDA) that existing methods adopt uniform alignment weights for all source-target view pairs, and proposes a Barlow Twins-guided adaptive alignment framework (BTGA). The method applies the invariance and redundancy reduction terms of Barlow Twins to each source-target view pair to quantify its consistency and redundancy, and dynamically weights the alignment strength accordingly. On 12 cross-domain node classification tasks across the ACM and MAG benchmarks, BTGA significantly outperforms existing baselines and achieves state-of-the-art performance.

Critical strengths

Identifies and exploits the empirical regularity that higher consistency and lower redundancy correspond to smaller domain discrepancy, converting Barlow Twins signals into adaptive weights for source–target view-pair alignment.

Critical weaknesses

1.Although the paper proposes using consistency and redundancy to dynamically assign view-pair alignment weights, the experiments did not compare the model to the same model using uniform weighting.

2.The correlation between the two statistics and domain discrepancy relies primarily on limited empirical observations (Fig. 1b), lacking sufficient theoretical justification and adequate experimental validation.

Questions for authors

1.The current ablation experiments lack comparisons with simple weighting strategies, such as uniform weighting. Could the authors consider adding these to support the performance gains of the proposed adaptive alignment mechanism?

2.This method assigns greater alignment weights to view pairs with smaller domain differences. Could the authors explain why these view pairs are prioritized for alignment over those with larger domain differences?

3.Figure 1 only illustrates the relationship between consistency, redundancy, and domain discrepancy on the US→CN task. Could the authors provide additional results on other transfer tasks to verify the generalizability of this observation?

4.Appendix B reports task-specific hyperparameter settings. Could the authors clarify how these hyperparameters are selected without using target-domain labels in the unsupervised domain adaptation setting? Even a brief theoretical explanation or empirical analysis would be appreciated.

6 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/2)

Thank you for your insightful and detailed feedback. We address your concerns point by point below.

W1 & Q1: Comparison with uniform weighting.

Thank you for suggesting a comparison with uniform weighting. We have added this comparison to the ablation study in Section 4.5 of the main text (p. 8–9, lines 431–438, Figure 4) to assess the contribution of adaptive view-pair weighting.

We use ordinary multi-view MMD alignment as the uniform-weight control, assigning equal weights to all source–target view pairs while keeping the other components unchanged. The table below reports some of the ACC results.

MethodACM2→ACM1US→CNJP→CNDE→CNFR→DEUS→RU
Ordinary multi-view MMD0.6410.5350.4930.4320.5360.460
BTGA0.6870.5580.5460.4610.5640.477

BTGA achieves higher ACC than uniform weighting on all twelve ACM and MAG tasks. These results support dynamically assigning alignment weights according to view-pair consistency and redundancy, enabling adaptive cross-domain alignment with better target classification. Full results with standard deviations are in Table E.4, Appendix E.4 (p. 25, lines 1327–1340).

W2 & Q3: Consistency, redundancy, and domain discrepancy.

Thank you for requesting stronger theoretical justification and broader experimental validation. We have expanded both aspects in Section 4.7 of the main text (p. 9–10, lines 463–495; Figure 6, p. 9, lines 483–485).

  • (i) Theoretical analysis. GAA [1] shows that the target-risk bound depends on node-attribute discrepancies ∥XiS−XjT∥22\|X_i^S-X_j^T\|_2^2 and structure-aggregated feature discrepancies ∥(ASXS)i−(ATXT)j∥22\|(A^SX^S)_i-(A^TX^T)_j\|_2^2, relating to attribute consistency and structure-related consistency, respectively. HGDA [2] likewise incorporates attribute and structural-signal discrepancies into its target-risk bound. Under the respective assumptions and with other terms fixed, reducing these discrepancies can tighten the risk bounds, supporting the prioritization of view pairs with better-matched attribute and structural information for transfer.

    BTGA's relation-view representations combine node attributes and relational structure, and CC measures cross-domain consistency at the feature level. Higher CC indicates stronger agreement between corresponding feature dimensions, providing a basis for comparing cross-domain view-pair matching.

    Corresponding-dimension agreement can coexist with repeated encoding of shared signals across dimensions. Inspired by the redundancy-reduction principle of Barlow Twins [3], BTGA uses RR to measure squared cross-domain correlations between different feature dimensions, capturing redundancy not reflected by CC. We believe that reducing redundant encoding while maintaining consistency in shared cross-domain information helps learn representations better suited to transfer and target classification. BTGA therefore jointly uses CC and RR to dynamically assign greater relative alignment weights to high-consistency, low-redundancy view pairs.

  • (ii) Additional experiments. We separately examine Consistency–MMD and Redundancy–MMD across all twelve ACM and MAG tasks. Figure 6 in Section 4.7 (p. 9, lines 483–485) presents three tasks. The table summarizes Pearson correlations using all observation points on each plotted four-pair mean curve.

    Relationship (Pearson r)ACM1→ACM2ACM2→ACM1US→CN
    Consistency–MMD−0.975−0.962−0.729
    Redundancy–MMD0.9930.9940.988

    The negative Consistency–MMD and positive Redundancy–MMD associations extend the original single-task observation, supporting the use of both statistics to characterize view-pair domain discrepancy. Together with the weighting comparison above, they provide empirical support for BTGA's consistency- and redundancy-guided adaptive alignment. Full experimental results are provided in Section 4.7 of the main text (p. 9–10, lines 463–495; Figure 6, p. 9, lines 483–485) and Appendix G.2 (p. 29, lines 1539–1549; Figure G.4, p. 30, lines 1596–1599).

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

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

[3] Zbontar et al. Barlow Twins: Self-Supervised Learning via Redundancy Reduction. ICML 2021.

AuthorsOpenReview ↗

Official Comment by Authors (2/2)

Q2: Prioritization of view-pair alignment.

Thank you for this question. We explain the design rationale and empirical evidence below, followed by a comparison with large-discrepancy prioritization. We have added the comparison to Section 4.5 of the main text (p.8–9, lines 431–438, Figure 4).

  • (i) Rationale and empirical evidence. Our aim is to prioritize information useful for target classification. We believe that smaller-discrepancy view pairs offer a more reliable basis for transfer. BTGA therefore gives them greater relative weights while retaining all pairs in alignment.

    We fix the task and target view, retaining the full model and other losses. From the same state after five epochs without alignment, we compare no alignment with aligning either source view (coefficient 1). The table below reports some of the results.

    Task (target view)Aligned pairMK-MMD²Target ACCNo-alignment ACC
    ACM2→ACM1 (PAP)PAP→PAP0.53450.5280.434
    ACM2→ACM1 (PAP)PSP→PAP0.80000.4710.434
    DE→CN (PP)PP→PP0.38810.4690.423
    DE→CN (PP)PAP→PP0.39180.4380.423

    Both alignment settings improve upon no alignment, but the smaller-discrepancy pair yields higher ACC by 5.7 and 3.1 percentage points, respectively. Thus, larger discrepancy does not necessarily imply greater alignment benefit. Full results are in Table G.3, Appendix G.3 (p.30, lines 1600–1610).

  • (ii) Comparison with large-discrepancy prioritization. In Section 4.5 of the main text (p.8–9, lines 431–438, Figure 4), we compare BTGA with a control assigning higher weights to larger-discrepancy view pairs, keeping other components unchanged. The table below reports some of the ACC results.

    MethodACM2→ACM1US→CNJP→CNDE→CNFR→DEUS→RU
    Large-discrepancy prioritization0.5660.4330.4920.4280.5210.424
    BTGA0.6870.5580.5460.4610.5640.477

    BTGA achieves higher ACC on all twelve tasks, supporting its strategy of dynamically assigning greater relative weights to high-consistency, low-redundancy view pairs. Full results with standard deviations are in Table E.4, Appendix E.4 (p.25, lines 1327–1340).

Q4: Hyperparameter selection without target-domain labels.

Thank you for highlighting this issue. We have clarified the original selection protocol and added supplementary analyses without target-domain labels in Appendix B.3 (p.17–20, lines 885–1051).

  • (i) A shared challenge in unsupervised domain adaptation. Hyperparameter and checkpoint selection is challenging without an independent labeled target-domain validation set. This challenge arises in both visual and graph domain adaptation.

  • (ii) Selection protocol for the main experiments. Following SAGA's emphasis on consistent training and evaluation [4], we compare BTGA and the baselines under the same data partitions and evaluation conditions. Hyperparameters are selected using target-domain test performance. Target labels are used for selection and evaluation, but not model parameter optimization.

  • (iii) Supplementary selection without target-domain labels. We have added Soft Neighborhood Density (SND) [5] and Importance Weighted Cross-Validation (IWCV) [6] analyses. SND maximizes a neighborhood-density score from unlabeled target predictions; IWCV minimizes an importance-weighted source-validation loss.

    We independently vary α\alpha and β\beta over {0.25,0.5,1,2,4}\{0.25,0.5,1,2,4\} times their task-specific defaults, keeping other hyperparameters fixed. Candidates are scored and selected using fixed final checkpoints, without target-domain labels. Representative results are:

    CriterionTaskParameterMultiplierSelected ACCMain ACC
    SNDDE→FRβ\beta0.5×0.5570.560
    IWCVCN→USα\alpha1×0.5880.592

    The selected configurations achieve ACC within 0.3 and 0.4 percentage points of the corresponding main results. These examples support target-label-free hyperparameter selection within the examined candidate ranges. Further selection analyses are provided in Figures B.2–B.3 (p.19–20, lines 1014–1045).

References

[4] Fang et al. SAGA: Structural aggregation guided alignment with dynamic view and neighborhood order selection for multiview graph domain adaptation. ICLR 2026.

[5] Saito et al. Tune it the right way: Unsupervised validation of domain adaptation via soft neighborhood density. ICCV 2021.

[6] Sugiyama et al. Covariate shift adaptation by importance weighted cross validation. JMLR 2007.

Thank you for the detailed response! The new comparisons address most of my concerns. My remaining question is about hyperparameter selection.

The revised manuscript states:

Specifically, all methods are trained for a predefined number of epochs, and their hyperparameters are selected according to target-domain test performance.

I understand that this follows SAGA and that the same selection rules apply to all methods. Still, target test labels influence which configuration is selected, even if they are not used to train the model parameters. The added SND analysis is a helpful step, but it varies only α and β around task-specific defaults. I would like to see whether BTGA’s advantage holds when the other hyperparameters are also chosen without target labels.

Could you compare BTGA with its uniform-weighting counterpart in a setting where target labels are used only for final evaluation? A small set of complete configurations would suffice. The configurations can be specified in advance or sampled from predefined ranges, without using target test performance to set the defaults or ranges. Please use the same SND criterion and tuning budget to select a configuration separately for each method, and fix the training duration in advance or use a label-free rule for checkpoint selection. This comparison would directly test whether adaptive weighting improves performance when the entire model selection process is independent of target labels.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

Dear Area Chair and Reviewers,

Thank you for the time and effort you have devoted to reviewing our paper. Your constructive comments have helped us strengthen the evidence and improve the presentation of our work. We appreciate your recognition of the following strengths:

  • Our paper addresses a relevant limitation of multi-view graph adaptation: domain shifts vary across relation pairs, motivating differentiated alignment (Reviewers 4sia, BK59, and 4QrY).

  • The central idea of weighting source–target view pairs by consistency and redundancy (C/R) is interpretable and well-motivated (Reviewer BK59). Reviewers pPmg and 4sia also recognize the use of view-pair statistics to guide adaptive alignment.

  • Our results consistently outperform the evaluated baselines across twelve transfer tasks, with particularly substantial improvements on ACM and gains in both ACC and Macro-F1 (Reviewer 4QrY). Reviewer BK59 also highlights the substantial and consistent gains on the ten MAG tasks.

  • Our component ablations help assess the proposed design (Reviewers 4sia and BK59), while parameter sensitivity analyses and computational comparisons provide useful evidence beyond classification results (Reviewer 4QrY). Reviewer BK59 also recognizes the method as computationally practical, with competitive runtime and the lowest peak GPU memory among the compared methods.

We have carefully considered your concerns and revised the manuscript accordingly. We have incorporated the key explanations and comparisons into the main text. Further details are provided in the appendices. The principal improvements are summarized below.

  • We appreciate your suggestions to better establish the contribution of adaptive weighting and the choice of alignment priority (Reviewers pPmg and 4QrY). We have added comparisons with uniform weighting and large-discrepancy prioritization. BTGA achieves higher ACC than both strategies on all twelve tasks, supporting adaptive alignment and the adopted prioritization strategy. The comparison is now included in Section 4.5 of the main text (p. 8–9, lines 431–438, Figure 4).

  • We value your suggestions to strengthen the justification for C/R-guided alignment (Reviewers pPmg, 4sia, BK59, and 4QrY). We have connected the weighting design with existing graph domain adaptation analyses and Barlow Twins (BT) principles, and expanded the C/R–MMD analysis from one task to all twelve. These additions strengthen the motivation and empirical basis for jointly using consistency and redundancy to evaluate cross-domain view pairs. The key explanations and results are now presented in Section 4.7 of the main text (p. 9–10, lines 463–495, Figure 6).

  • In response to your questions about the different roles of BT within and across domains (Reviewers 4sia and BK59), we have clarified why BT regularizes representations within domains while its statistics guide cross-domain alignment. We have also added two replacement comparisons: direct cross-domain BT and intra-domain MK-MMD. Full BTGA achieves higher ACC and Macro-F1 than both replacements on all twelve tasks, supporting the adopted division of roles. This rationale and comparison are included in Section 4.7 of the main text (p. 10, lines 496–504).

  • Addressing your questions on hyperparameter and checkpoint selection (Reviewers pPmg and 4QrY), we have revised the parameter analysis in Section 4.8 of the main text (p. 10, lines 506–529, Figure 7) and clarified the original selection protocol. We have also added supplementary SND and IWCV analyses, in which candidates are scored and selected without target-domain labels using fixed final checkpoints (Appendix B.3, p. 17–20, lines 885–1051).

  • Following your suggestion on statistical reporting (Reviewer 4QrY), we now report mean ± standard deviation across five runs for the main BTGA experiments, component ablations, and filtering/alignment comparisons. These additions allow the variability to be assessed alongside the reported performance (Tables E.1–E.4, p. 24–25, lines 1273–1340).

To support independent reproduction, we will make our source code publicly available upon acceptance.

We hope these revisions and our detailed responses under each review clarify the design choices and strengthen the evidence for BTGA. Thank you again for your time and constructive suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.