acceptodds
Under review as a conference paper at ICLR 2027

CARE: Cross-Domain Alignment by Reweighting Edges 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, where nodes are connected through multiple relation views. Existing methods typically perform message passing over fixed edge weights, so information specific to a domain or a relation view is aggregated layer after layer along each view's connection pattern and carried into the node representations that alignment must then reconcile. We propose Cross-domain Alignment By Reweighting Edges (CARE), an edge-reweighting framework for MGDA. CARE updates the edge weights of each relation view under the cross-domain alignment objective: edges that propagate information shared across domains are strengthened, and edges that propagate domain-specific or view-specific information are suppressed. 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 3.33 over 3 reviews, above 96% of the others.
Reviewer_bZLnOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes CARE, an edge-reweighting framework for multi-view graph domain adaptation. CARE uses endpoint features and Gumbel-Sigmoid reparameterization to learn continuous weights for the edges of each relation view in both the source and target domains. The learned weights are bounded within [1−ρ,1+ρ][1-\rho,1+\rho]. APPNP propagation is then performed on the reweighted graphs, followed by a shared MLP encoder.

Critical strengths
  1. The research question is interesting and valuable.

  2. The paper is well-written, organized, and easy to follow

  3. To the best of my knowledge, this is a novel approach.

  4. Using the downstream alignment objective to guide message-passing edge weights is more adaptive than propagating over a fixed graph.

  5. Bounding the learned edge weights within [1−ρ,1+ρ][1-\rho,1+\rho] prevents complete edge removal or uncontrolled amplification.

In summary, I do believe that this work is interesting! By the way, all figures are beautiful! My eyes enjoy it so much.

Critical weaknesses
  1. Eq. (9) only maximizes the log-sum-exp of cross-domain cosine similarities. It contains no negative pairs, class-conditional constraints, variance regularization, or one-to-one matching. Consider pu,iS=pv,jT=cfor all i,jp^S_{u,i}=p^T_{v,j}=c \quad \text{for all } i,j, where cc is any nonzero vector. All cosine similarities then equal one, attaining the global optimum of the objective. Because classification is performed using Z~\widetilde Z, whereas alignment is performed through separate projection heads gvg_v, the source classification loss does not necessarily prevent the projection heads from collapsing.

  2. Figure 1 uses ground-truth labels to measure class-pair edge-distribution discrepancies. These target labels are unavailable during training. CARE neither estimates class-pair edge distributions nor explicitly minimizes their discrepancy; it only learns edge weights indirectly through a generic representation-similarity objective.

  3. The edge score is computed as MLP⁡ ⁣([WXi;WXj])\operatorname{MLP}\!\left([WX_i;WX_j]\right), which is generally not invariant to swapping ii and jj. Therefore, (i,j)(i,j) and (j,i)(j,i) may receive different weights. This is inconsistent with the undirected-graph assumption and symmetric normalization unless the weights are explicitly symmetrized.

Questions for authors
  1. How does the method avoid the collapsed solution of Eq. (9), in which every projected source and target representation is identical?

  2. Can the authors report the learned edge-weight distributions and quantify their relationship with edge transferability?

  3. Do the class-pair edge-distribution discrepancies shown in Figure 1 decrease after edge reweighting?

  4. Is the constraint ω~i,j=ω~j,i\widetilde\omega_{i,j}=\widetilde\omega_{j,i} explicitly enforced for undirected edges?

  5. Why is the closest structural reweighting graph domain adaptation method not included as a baseline?

  6. Is Gumbel noise still sampled during inference? If so, how is prediction stability ensured?

11 comments
Reviewer_bZLnOpenReview ↗

ICLR Rebuttal: Please Submit Your Response Soon

This is a reminder that the ICLR rebuttal period is ending soon. I would appreciate it if the authors could submit their response as soon as possible so that I can finalize my review.

AuthorsOpenReview ↗

Official Comment by Authors (1/6)

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

Response to W1: Projection Representation Collapse in Eq. (9)

We thank you for raising the question of potential projection representation collapse in the alignment objective. We analyze the learned projected representations after training to verify that projection collapse does not occur in CARE.

Class-Conditional Alignment Analysis. To examine whether different classes remain distinguishable after cross-domain alignment, we compute the class centers of source and target projected representations after training. For the vv-th view of domain D∈S,TD\in\\{S,T\\}, let Nv,cD=i:yiD=c\mathcal N_{v,c}^D=\\{i:y_i^D=c\\} denote the set of nodes belonging to class cc, and Nv,cDN_{v,c}^D denote the corresponding number of nodes. The class center is defined as:

μv,cD=1Nv,cD∑i∈Nv,cDpv,iD\boldsymbol{\mu}_{v,c}^D=\frac{1}{N_{v,c}^D}\sum_{i\in\mathcal N_{v,c}^D}\mathbf p_{v,i}^D

Here, pv,iD\mathbf p_{v,i}^D denotes the projected representation of node ii in the vv-th view. For source class cc and target class kk, the absolute and relative distances are defined as:

d_{ck}=\left\\|\boldsymbol{\mu}_{v,c}^S-\boldsymbol{\mu}_{v,k}^T\right\\|_2,\qquad D_{ck}^{\mathrm{rel}}=\frac{d_{ck}}{d_{cc}}

Here, μv,cS\boldsymbol{\mu}_{v,c}^S and μv,kT\boldsymbol{\mu}_{v,k}^T denote the projected class centers of source class cc and target class kk in the vv-th view, respectively. The diagonal entries of the relative distance matrix are 1, while the off-diagonal entries represent the distances between different-class centers relative to the corresponding same-class center distances.

Taking CN→JP as an example, the relative distances between projected representations of different classes are shown in Table 1.

Table 1. Relative class-center distances of CARE (CN-JP, PAP-PAP).

S/TC0C1C2C3C4C5C6C7C8C9
C01.007.585.368.085.705.149.816.082.892.40
C13.401.002.805.133.113.825.902.983.963.68
C23.784.561.006.152.882.397.571.804.354.31
C311.2015.7012.051.0011.3910.3915.5211.679.9011.21
C42.083.742.404.071.000.755.862.323.283.47
C54.064.492.846.361.511.007.962.734.514.76
C66.008.987.182.276.056.091.006.956.346.59
C73.103.361.974.302.151.945.341.003.353.51
C82.016.674.885.124.724.326.515.181.001.59
C90.474.383.104.663.242.905.693.461.431.00

The relative distances of projected representations for other transfer tasks are provided in Tables 3–10 of Appendix H. The results show that, although Eq. (9) does not explicitly impose class-level constraints, the vast majority of same-class center distances are substantially smaller than different-class center distances, indicating that CARE's projected representations retain class-discriminative structure. This may be attributed to objectives such as source-domain classification, which encourage the encoder to preserve discriminative information and provide class-structured inputs to the projection heads.

AuthorsOpenReview ↗

Official Comment by Authors (2/6)

Response to Q1: Projection Representation Collapse in Eq. (9)

We thank you for raising the concern about projection representation collapse. To verify that projection collapse does not occur in CARE, we compute the class-center distances between source and target projected representations after training, as reported in Response to W1.

However, the alignment objective in Eq. (9) neither explicitly considers class information nor theoretically rules out trivial collapsed solutions. To further enhance class discrimination and reduce collapse risk, we introduce class-conditional alignment and variance regularization.

Class-Conditional Alignment and Variance Regularization. We first compute the source and target class centers for the vv-th view:

μv,cD=1Nv,cD∑i∈Nv,cDpv,iD\boldsymbol{\mu}_{v,c}^D=\frac{1}{N_{v,c}^D}\sum_{i\in\mathcal N_{v,c}^D}\mathbf p_{v,i}^D

Here, pv,iD\mathbf p_{v,i}^D denotes the projected representation of node ii in the vv-th view of domain D∈S,TD\in\\{S,T\\}, Nv,cS=i:yiS=c\mathcal N_{v,c}^S=\\{i:y_i^S=c\\} contains source nodes of class cc, and Nv,cT\mathcal N_{v,c}^T contains target nodes with high-confidence pseudo-labels of class cc. Nv,cDN_{v,c}^D denotes the corresponding number of nodes. We then employ symmetric cross-entropy to encourage cross-domain same-class matching and different-class discrimination:

Qck=(μv,cS)⊤μv,kTτ,Lcls=12[LCE(Q,I)+LCE(Q⊤,I)]Q_{ck}=\frac{(\boldsymbol{\mu}_{v,c}^S)^\top\boldsymbol{\mu}_{v,k}^T}{\tau},\qquad \mathcal L_{\mathrm{cls}}=\frac12\left[\mathcal L_{\mathrm{CE}}(Q,I)+\mathcal L_{\mathrm{CE}}(Q^\top,I)\right]

Here, II denotes the identity matching matrix. To further prevent excessive concentration of projected representations, we introduce variance regularization that penalizes feature dimensions with standard deviations below a threshold:

Lvar=12d∑D∈S,T∑j=1dmax⁡(0,mvar−σv,jD)\mathcal L_{\mathrm{var}}=\frac{1}{2d}\sum_{D\in\\{S,T\\}}\sum_{j=1}^{d}\max(0,m_{\mathrm{var}}-\sigma_{v,j}^D)

σv,jD=1NvD∑i=1NvD(ξv,i,jD−ξˉv,jD)2+ε\sigma_{v,j}^D=\sqrt{\frac{1}{N_v^D}\sum_{i=1}^{N_v^D}(\xi_{v,i,j}^D-\bar{\xi}_{v,j}^D)^2+\varepsilon}

Here, ξv,i,jD=[pv,iD]j\xi_{v,i,j}^D=[\mathbf p_{v,i}^D]_j denotes the jj-th component of node ii's projected representation in the vv-th view, ξˉv,jD=1NvD∑i=1NvDξv,i,jD\bar{\xi}_{v,j}^D=\frac{1}{N_v^D}\sum_{i=1}^{N_v^D}\xi_{v,i,j}^D is its mean, NvDN_v^D is the number of nodes used for computation in the vv-th view of domain DD, dd is the projection dimension, mvarm_{\mathrm{var}} is the standard deviation threshold, and ε\varepsilon ensures numerical stability. By penalizing low-variance dimensions, this regularization reduces the risk of projection collapse.

We incorporate both losses into the original CARE objective and denote the resulting variant as CARE-1. Taking CN→JP as an example, we further evaluate its relative class-center distances and target-domain classification performance.

Table 1. Relative class-center distances of CARE-1 (CN-JP, PAP-PAP).

S/TC0C1C2C3C4C5C6C7C8C9
C01.0041.9324.1923.4122.5922.6729.3322.613.651.80
C130.811.0014.5715.4215.9815.9111.5415.7029.0830.31
C274.0859.161.0013.189.128.9521.236.1966.1871.80
C332.3027.025.801.005.875.797.456.2028.7331.32
C413.7411.341.412.291.000.943.971.8012.2513.31
C522.8619.952.443.821.121.007.272.7220.3222.13
C643.1723.5010.008.4711.7811.641.0011.9239.2342.07
C721.4718.380.933.942.642.617.001.0019.1120.79
C84.1746.2524.3723.3522.3922.4830.6122.461.003.26
C90.2748.8327.6426.7225.7325.8233.7825.753.181.00

Compared with the original CARE results in Response to W1, most different-class centers exhibit larger relative distances, indicating that incorporating class information into the alignment objective improves the discrimination between different classes.

Downstream Performance.

TaskACC (CARE)ACC (CARE-1)Macro-F1 (CARE)Macro-F1 (CARE-1)
CN-US0.5930.5970.4200.426
JP-CN0.5320.5310.4020.400
CN-JP0.5770.5820.4460.451
DE-FR0.5430.5480.4410.444
RU-US0.3890.3890.3870.386
ACM1-ACM20.5380.5430.4320.437

Overall, target-domain classification performance remains stable after introducing the additional losses. However, pseudo-label filtering and class-center computation may incur additional computational overhead. Therefore, we retain the simpler original objective. We thank you for highlighting this issue, which has prompted us to examine it more thoroughly.

Reviewer_u5fPOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CARE for multi-view graph domain adaptation. CARE learns edge weights in both source and target graphs, using cross-domain alignment to guide message propagation in each relation view. The model combines corresponding-view and cross-view alignment with source classification and target entropy minimization. Experiments on ACM and MAG show improvements over the compared baselines across all 12 transfer tasks.

Critical strengths
  1. The method directly connects cross-domain alignment with edge reweighting. This provides a reasonable way to adapt message propagation to domain shifts across relation views, while allowing both source and target structures to be adjusted.

  2. The reported improvements are consistent across the evaluated tasks. The ablation that blocks alignment gradients to the edge-weighting module is particularly useful, as it tests whether alignment-guided reweighting contributes beyond representation alignment alone.

Critical weaknesses
  1. The alignment objective in Eq. (9) increases aggregate similarity between source and target representations without distinguishing classes. This raises a concern about aligning nodes from different classes. The source classification loss provides some constraint, but the paper does not establish whether the learned edge weights actually preserve transferable class information. An analysis of the weighted edges and class-conditional alignment would help support the proposed mechanism.

  2. The model selection protocol is unclear. The appendix reports task-specific hyperparameters but does not explain how they are selected, or how training is stopped, without target labels.

Questions for authors

See weaknesses

6 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/4)

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

Response to W1: Class-Conditional Alignment and Edge Weight Analysis

We thank you for raising questions about class-conditional alignment and edge weight analysis. To further evaluate whether CARE preserves class-discriminative information while promoting cross-domain representation alignment, we conduct additional class-conditional alignment and edge weight analyses.

Class-Conditional Alignment Analysis. To examine whether different classes remain distinguishable after cross-domain alignment, we compute the class centers of source and target projected representations after training. For the vv-th view of domain D∈S,TD\in\\{S,T\\}, let Nv,cD=i:yiD=c\mathcal N_{v,c}^D=\\{i:y_i^D=c\\} denote the set of nodes belonging to class cc, and Nv,cDN_{v,c}^D denote the corresponding number of nodes. The class center is defined as:

μv,cD=1Nv,cD∑i∈Nv,cDpv,iD\boldsymbol{\mu}_{v,c}^D=\frac{1}{N_{v,c}^D}\sum_{i\in\mathcal N_{v,c}^D}\mathbf p_{v,i}^D

Here, pv,iD\mathbf p_{v,i}^D denotes the projected representation of node ii in the vv-th view. For source class cc and target class kk, the absolute and relative distances are defined as:

d_{ck}=\left\\|\boldsymbol{\mu}_{v,c}^S-\boldsymbol{\mu}_{v,k}^T\right\\|_2,\qquad D_{ck}^{\mathrm{rel}}=\frac{d_{ck}}{d_{cc}}

Here, μv,cS\boldsymbol{\mu}_{v,c}^S and μv,kT\boldsymbol{\mu}_{v,k}^T denote the projected class centers of source class cc and target class kk in the vv-th view, respectively. The diagonal entries of the relative distance matrix are 1, while the off-diagonal entries represent the distances between different-class centers relative to the corresponding same-class center distances.

Taking CN→JP as an example, the relative distances between projected representations of different classes are shown in Table 1.

Table 1. Relative class-center distances of CARE (CN-JP, PAP-PAP).

S/TC0C1C2C3C4C5C6C7C8C9
C01.007.585.368.085.705.149.816.082.892.40
C13.401.002.805.133.113.825.902.983.963.68
C23.784.561.006.152.892.397.571.804.354.31
C311.2015.7012.051.0011.3910.3915.5211.679.9011.21
C42.083.742.404.071.000.755.862.323.283.47
C54.064.492.846.361.521.007.962.734.514.76
C66.008.987.182.276.056.091.006.956.346.59
C73.103.361.974.302.151.945.351.003.353.51
C82.016.684.895.124.724.326.525.181.001.59
C90.474.383.104.663.242.905.693.461.431.00

The relative distances of projected representations for other transfer tasks are provided in Table 3-10 of Appendix H. The results show that, although Eq. (9) does not explicitly impose class-level constraints, the vast majority of same-class center distances are substantially smaller than different-class center distances, indicating that CARE's projected representations retain class-discriminative structure. This may be attributed to objectives such as source-domain classification, which encourage the encoder to preserve discriminative information and provide class-structured inputs to the projection heads.

AuthorsOpenReview ↗

Official Comment by Authors (2/4)

Response to W1: Class-Conditional Alignment and Edge Weight Interpretation

We thank you for raising questions about class-conditional alignment and edge weight analysis. As demonstrated in Response to W1, the original CARE retains class-discriminative structure without projection collapse.

However, the alignment objective in Eq. (9) does not explicitly consider class information or theoretically rule out trivial collapsed solutions. To further enhance class discrimination and reduce collapse risk, we introduce class-conditional alignment and variance regularization.

Class-Conditional Alignment and Variance Regularization. We first compute the source and target class centers for the vv-th view:

μv,cD=1Nv,cD∑i∈Nv,cDpv,iD\boldsymbol{\mu}_{v,c}^D=\frac{1}{N_{v,c}^D}\sum_{i\in\mathcal N_{v,c}^D}\mathbf p_{v,i}^D

Here, pv,iD\mathbf p_{v,i}^D denotes the projected representation of node ii in the vv-th view of domain D∈S,TD\in\\{S,T\\}, Nv,cS=i:yiS=c\mathcal N_{v,c}^S=\\{i:y_i^S=c\\} contains source nodes of class cc, and Nv,cT\mathcal N_{v,c}^T contains target nodes with high-confidence pseudo-labels of class cc. Nv,cDN_{v,c}^D denotes the corresponding number of nodes. We then employ symmetric cross-entropy to encourage cross-domain same-class matching and different-class discrimination:

Qck=(μv,cS)⊤μv,kTτ,Lcls=12[LCE(Q,I)+LCE(Q⊤,I)]Q_{ck}=\frac{(\boldsymbol{\mu}_{v,c}^S)^\top\boldsymbol{\mu}_{v,k}^T}{\tau},\qquad \mathcal L_{\mathrm{cls}}=\frac12\left[\mathcal L_{\mathrm{CE}}(Q,I)+\mathcal L_{\mathrm{CE}}(Q^\top,I)\right]

Here, II denotes the identity matching matrix. To further prevent excessive concentration of projected representations, we introduce variance regularization that penalizes feature dimensions with standard deviations below a threshold:

Lvar=12d∑D∈S,T∑j=1dmax⁡(0,mvar−σv,jD)\mathcal L_{\mathrm{var}}=\frac{1}{2d}\sum_{D\in\\{S,T\\}}\sum_{j=1}^{d}\max(0,m_{\mathrm{var}}-\sigma_{v,j}^D)

σv,jD=1NvD∑i=1NvD(ξv,i,jD−ξˉv,jD)2+ε\sigma_{v,j}^D=\sqrt{\frac{1}{N_v^D}\sum_{i=1}^{N_v^D}(\xi_{v,i,j}^D-\bar{\xi}_{v,j}^D)^2+\varepsilon}

Here, ξv,i,jD=[pv,iD]j\xi_{v,i,j}^D=[\mathbf p_{v,i}^D]_j denotes the jj-th component of node ii's projected representation in the vv-th view, ξˉv,jD=1NvD∑i=1NvDξv,i,jD\bar{\xi}_{v,j}^D=\frac{1}{N_v^D}\sum_{i=1}^{N_v^D}\xi_{v,i,j}^D is its mean, NvDN_v^D is the number of nodes used for computation in the vv-th view of domain DD, dd is the projection dimension, mvarm_{\mathrm{var}} is the standard deviation threshold, and ε\varepsilon ensures numerical stability. By penalizing low-variance dimensions, this regularization reduces the risk of projection collapse.

We incorporate both losses into the original CARE objective and denote the resulting variant as CARE-1. Taking CN→JP as an example, we further evaluate its relative class-center distances and target-domain classification performance.

Table 1. Relative class-center distances of CARE-1 (CN-JP, PAP-PAP).

S/TC0C1C2C3C4C5C6C7C8C9
C01.0041.9324.1923.4122.5922.6729.3322.613.651.80
C130.811.0014.5715.4215.9815.9111.5415.7029.0830.31
C274.0859.161.0013.189.128.9521.236.1966.1871.80
C332.3027.025.801.005.875.797.456.2028.7331.32
C413.7411.341.412.291.000.943.971.8012.2513.31
C522.8619.952.443.821.121.007.272.7220.3222.13
C643.1723.5010.008.4711.7811.641.0011.9239.2342.07
C721.4718.380.933.942.642.617.001.0019.1120.79
C84.1746.2524.3723.3522.3922.4830.6122.461.003.26
C90.2748.8327.6426.7225.7325.8233.7825.753.181.00

Compared with the original CARE results in Response to W1, most different-class centers exhibit larger relative distances, indicating that incorporating class information into the alignment objective improves class discrimination.

Downstream Performance.

TaskACC (CARE)ACC (CARE-1)Macro-F1 (CARE)Macro-F1 (CARE-1)
CN-US0.5930.5970.4200.426
JP-CN0.5320.5310.4020.400
CN-JP0.5770.5820.4460.451
DE-FR0.5430.5480.4410.444
RU-US0.3890.3890.3870.386
ACM1-ACM20.5380.5430.4320.437

Overall, target-domain classification performance remains stable after introducing the additional losses. However, pseudo-label filtering and class-center computation may incur additional computational overhead. We therefore retain the simpler original objective. We thank you for highlighting this issue, which has prompted us to examine it more thoroughly.

AuthorsOpenReview ↗

Official Comment by Authors (3/4)

Response to W1: Class-Conditional Alignment and Edge Weight Interpretation

We thank you for raising questions about class-conditional alignment and edge weight analysis. To further evaluate whether CARE preserves transferable class-discriminative information, we conduct two additional post-hoc experiments, edge pruning and domain discriminability analysis, to examine the roles of edges with different learned weights in downstream classification and cross-domain representation discriminability.

(1) Edge Pruning Analysis. We use the pretrained CARE generator to compute and rank edge weights, then remove the top kk% highest-weight edges (High), the bottom kk% lowest-weight edges (Low), or a randomly selected kk% of edges (Random). After pruning, we fix the original learned weights of the remaining edges and retrain the downstream graph domain adaptation model.

Taking ACM1→ACM2 as an example, the target-domain Macro-F1 under different edge deletion ratios is shown in Table 1.

Table 1. Target-domain Macro-F1.

Edge Deletion Ratio (%)High (%)Low (%)Random (%)
043.0043.0043.00
1042.1543.0542.92
2041.8242.9942.78
3041.4842.9742.81
4041.3542.9242.52
5040.9542.8842.35

Complete results for other transfer tasks are provided in Figure 3 of the Experiments section and Figure 4 of Appendix I. As the proportion of removed high-weight edges increases, target-domain Macro-F1 declines markedly, indicating that high-weight edges facilitate the propagation of cross-domain shared information and improve transfer performance. In contrast, removing low-weight edges has a smaller impact and even slightly improves performance at certain ratios, indicating that these edges may propagate more domain-specific information that interferes with transferable information propagation and reduces transfer performance.

(2) Domain Discriminability Analysis. We further train a binary domain classifier on source and target node representations under different edge deletion ratios and use domain classification AUC to measure domain discriminability. A higher AUC indicates that source and target representations are more distinguishable, whereas a lower AUC indicates reduced domain discriminability.

Taking JP→CN as an example, domain classification AUC under different edge deletion ratios is shown in Table 2.

Table 2. Domain classification AUC.

Edge Deletion Ratio (%)High (%)Low (%)Random (%)
073.3773.3773.37
1073.9072.5073.10
2074.2571.9073.08
3074.3271.8872.90
4074.2871.7072.80
5074.3371.5572.70

Complete results for other transfer tasks are provided in Figure 4 of the Experiments section. Removing high-weight edges generally increases AUC, indicating that their removal weakens the propagation of cross-domain shared information, enlarges representation discrepancies between domains, and makes the domains more distinguishable. In contrast, removing low-weight edges generally decreases AUC, indicating that their removal reduces the propagation of domain-specific information, narrows cross-domain representation discrepancies, and makes the domains less distinguishable.

In summary, removing high-weight edges generally reduces downstream classification performance while increasing domain discriminability, whereas removing low-weight edges has a smaller impact on classification and decreases domain discriminability. These experiments provide complementary post-hoc empirical evidence for the relationship between learned edge weights and cross-domain transferable information, from the perspectives of downstream task performance and cross-domain representation distributions.

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

This paper studies multi-view graph domain adaptation, which transfers knowledge from a labeled source graph to an unlabeled target graph. To address the concern that message passing over fixed edge weights may repeatedly aggregate information specific to a domain or a relation view, the paper proposes CARE, which uses cross-domain alignment to guide updates to the edge weights of each relation view and improve target-domain node classification.

Critical strengths

CARE combines cross-domain representation alignment with adaptive edge reweighting, allowing gradients from the alignment loss to be backpropagated through graph propagation to update edge weights in each relation view of both the source and target domains. Experimental results show that CARE outperforms the compared baseline methods on multiple cross-network node classification tasks. Targeted ablation studies further support the contribution of this design to classification performance.

Critical weaknesses
  1. Insufficient validation of the motivation. The paper shows differences in edge distributions and lower performance for the variant using fixed edge weights. However, it does not directly demonstrate how these differences, through message passing over fixed edge weights, interfere with alignment and reduce transfer performance.

  2. Lack of direct evidence for the interpretation of edge weights. The reported performance and ablation results do not establish that the gains indeed arise from strengthening edges that propagate information shared across domains and suppressing edges that propagate domain-specific information, as the authors claim. This interpretation requires validation at the level of individual edges and cannot be inferred from classification improvements alone.

Questions for authors
  1. Could the authors hold node features and the proportion of nodes in each class fixed and directly examine how differences in connectivity affect propagation, alignment, and classification performance?

  2. Could the authors validate the above interpretation through an analysis of edge weights or experiments that modify edges? Without such evidence, I suggest explicitly presenting this interpretation as a design intention rather than a verified conclusion.

4 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/2)

Response to W1 & Q1: Validation of the Motivation

We thank you for this important suggestion. To further evaluate how cross-domain connectivity differences affect representation alignment and transfer performance through message passing with fixed edge weights, we conduct additional controlled experiments on synthetic graphs.

Experimental Setup. We construct source and target graphs using a four-class contextual stochastic block model (CSBM), each containing 6,000 nodes. We fix the source graph structure, node features, class proportions, and edge weights, and modify only the target-domain connectivity through constrained edge swaps while preserving node degrees and homophily. By varying the edge swap ratio ϵ∈0,0.25,0.50,0.75,1.00\epsilon\in\\{0,0.25,0.50,0.75,1.00\\}, we introduce different degrees of cross-domain structural shifts.

Evaluation Metrics. To evaluate the effects of structural shifts on message passing and representation alignment, we compute cross-domain class-center distances in the propagation and projection spaces. For domain D∈S,TD\in\\{S,T\\}, let NcD=i:yiD=c\mathcal N_c^D=\\{i:y_i^D=c\\} denote the set of nodes belonging to class cc, and NcDN_c^D denote the corresponding number of nodes. The class centers in the two spaces are defined as:

νcD=1NcD∑i∈NcDziD,μcD=1NcD∑i∈NcDpiD\boldsymbol{\nu}_c^D=\frac{1}{N_c^D}\sum_{i\in\mathcal N_c^D}\mathbf z_i^D,\qquad \boldsymbol{\mu}_c^D=\frac{1}{N_c^D}\sum_{i\in\mathcal N_c^D}\mathbf p_i^D

Here, ziD\mathbf z_i^D denotes the node representation obtained through graph propagation and encoding, while piD\mathbf p_i^D denotes the representation obtained through the projection head. Ground-truth target labels are used only for post-training class-center computation, not for model training.

Since graph propagation aggregates neighborhood information based on connectivity, cross-domain connectivity differences may shift the propagated representations of same-class nodes. We therefore use cross-domain same-class center distances in the propagation and projection spaces to measure the effects of connectivity changes on message passing and cross-domain alignment, respectively:

D_{\mathrm{prop}}=\frac{1}{C}\sum_{c=1}^{C}\left\\|\boldsymbol{\nu}_c^S-\boldsymbol{\nu}_c^T\right\\|_2

D_{\mathrm{align}}=\frac{1}{C}\sum_{c=1}^{C}\left\\|\boldsymbol{\mu}_c^S-\boldsymbol{\mu}_c^T\right\\|_2

Here, C=4C=4 is the number of classes. DpropD_{\mathrm{prop}} measures cross-domain discrepancies in node representations obtained through graph propagation and encoding with fixed edge weights, while DalignD_{\mathrm{align}} measures cross-domain same-class distances in the projection space after representation learning and alignment. These metrics reflect the effects of connectivity changes on message passing and cross-domain alignment, respectively.

Results and Analysis. We progressively increase the edge swap ratio to enlarge cross-domain edge distribution discrepancies. Table 1 reports the average results over five random seeds. Full experimental details are provided in Appendix G.

Table 1. Effects of connectivity shifts on propagation, alignment, and target classification.

Edge Swap RatioDpropD_{\mathrm{prop}}DalignD_{\mathrm{align}}Target Macro-F1
0.000.0710.0500.958
0.250.0950.0830.953
0.500.1650.1450.937
0.750.2400.1950.918
1.000.3150.2250.894

(1) Cross-domain connectivity changes affect message passing. As cross-domain edge distribution discrepancies increase, DpropD_{\mathrm{prop}} increases from approximately 0.071 to 0.315. With node features, class proportions, node degrees, and homophily held constant, connectivity changes alter neighborhood information aggregation, leading to larger discrepancies between same-class node representations.

(2) Cross-domain connectivity changes affect alignment. As cross-domain edge distribution discrepancies increase, DalignD_{\mathrm{align}} increases from approximately 0.050 to 0.225. This indicates that larger connectivity shifts increase the discrepancies between projected representations of same-class nodes across domains, making cross-domain alignment more challenging.

(3) Structural shifts further affect transfer performance. Correspondingly, the target Macro-F1 decreases from 0.958 to 0.894, a drop of approximately 6.4 percentage points. This shows that connectivity changes affect not only message passing and cross-domain alignment but also target-domain node classification performance.

In summary, with node features, class proportions, and graph homophily held constant, these experiments reveal the progressive effects of cross-domain connectivity changes on message passing, representation alignment, and downstream classification performance. These findings provide direct empirical support for CARE's motivation.

AuthorsOpenReview ↗

Official Comment by Authors (2/2)

Response to W2 & Q2:Interpretation of Learned Edge Weights

We thank you for raising this important question. To further evaluate the relationship between learned edge weights and cross-domain transferable information, we conduct two post-hoc analyses, edge pruning and domain classification, to examine the roles of edges with different weights from the perspectives of downstream classification performance and cross-domain representation discriminability.

(1) Edge Pruning Analysis. We use the pretrained CARE generator to compute and rank edge weights, then remove the top kk% highest-weight edges (High), the bottom kk% lowest-weight edges (Low), or randomly selected kk% edges (Random). After pruning at different ratios, we retain the original weights of the remaining edges and retrain the downstream graph domain adaptation model.

Taking ACM1→ACM2 as an example, Table 1 reports the target-domain Macro-F1 under different edge deletion ratios.

Table 1. Target-domain Macro-F1.

Edge Deletion Ratio (%)High (%)Low (%)Random (%)
043.0043.0043.00
1042.1543.0542.92
2041.8242.9942.78
3041.4842.9742.81
4041.3542.9242.52
5040.9542.8842.35

Complete results for other transfer tasks are provided in Figure 3 of the Experiments section and Figure 4 of Appendix I. As the proportion of removed high-weight edges increases, the target Macro-F1 declines markedly, indicating that these edges better facilitate the propagation of cross-domain shared information and improve transfer performance. In contrast, removing low-weight edges has little impact on performance and even leads to slight improvements at certain ratios, indicating that these edges tend to propagate more domain-specific information, interfering with shared information propagation and degrading transfer performance.

(2) Domain Discriminability Analysis. We further train a binary domain classifier on source and target node representations under different edge deletion ratios, using domain classification AUC to measure cross-domain discriminability. A higher AUC indicates that the two domains are more distinguishable, while a lower AUC indicates reduced discriminability.

Taking JP→CN as an example, Table 2 reports the domain classification AUC under different edge deletion ratios.

Table 2. Domain classification AUC.

Edge Deletion Ratio (%)High (%)Low (%)Random (%)
073.3773.3773.37
1073.9072.5073.10
2074.2571.9073.08
3074.3271.8872.90
4074.2871.7072.80
5074.3371.5572.70

Complete results for other transfer tasks are provided in Figure 4 of the Experiments section. Removing high-weight edges generally increases AUC, indicating that their removal disrupts the propagation of cross-domain shared information, enlarges domain discrepancies, and makes the domains more distinguishable. In contrast, removing low-weight edges generally decreases AUC, indicating that their removal reduces the propagation of domain-specific information, narrows domain discrepancies, and makes the domains less distinguishable.

In summary, removing high-weight edges generally degrades downstream classification performance while increasing domain discriminability, whereas removing low-weight edges has limited impact on classification performance and reduces domain discriminability. These two analyses provide complementary post-hoc empirical evidence linking learned edge weights to cross-domain transferable information from the perspectives of downstream task performance and cross-domain representation distributions.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer bUsc,

Thank you again for your time, effort, and constructive feedback. In response to your concerns, we have conducted additional experiments to validate our research motivation and the interpretation of learned edge weights. Your valuable suggestions have helped us substantially improve the quality of our paper.

As the discussion deadline approaches, we would greatly appreciate any further feedback you may have. If there are any remaining concerns, we will do our best to address them.

Best regards,

The Authors of Submission 31124.

Thank you for the detailed response and the additional experiments. I will raise my score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

Global Response

Dear reviewers,

Thank you for dedicating your valuable time and effort to reviewing our paper. We appreciate your insightful feedback and constructive comments, which have greatly helped us enhance the presentation of our work. We are grateful for your recognition of the contributions and strengths of our paper, as outlined below:

  • Our work combines cross-domain representation alignment with adaptive edge reweighting, using cross-domain alignment objectives to guide message propagation across different relation views in the source and target domains. The research problem is interesting and valuable, and the proposed approach offers a novel perspective on addressing multi-view graph domain adaptation. (Reviewers bUSc, u5fP, and bZLn)

  • Extensive experiments demonstrate the effectiveness of CARE, which consistently outperforms the compared MGDA baselines across multiple cross-domain node classification tasks. The ablation studies further show that alignment-guided edge reweighting provides additional benefits beyond representation alignment alone, demonstrating greater adaptability than propagation over fixed graph structures. (Reviewers bUSc, u5fP, and bZLn)

  • The paper is well-written, clearly structured, and logically organized. The presentation of the research and methodological design is easy to follow, helping readers understand the main ideas and contributions of the work. (Reviewer bZLn)

We have carefully considered each of your questions, concerns, and comments. In preparing the revised manuscript, we have addressed the issues raised and made corresponding revisions to improve the clarity and rigor of our work. The main changes are summarized below:

  • We thank you for your suggestion to further validate the research motivation through additional experiments (Reviewer bUSc). In response, we designed controlled experiments on synthetic CSBM graphs by modifying edge connections to examine their effects on message propagation discrepancy, representation alignment discrepancy, and target-domain node classification performance (see Appendix G).

  • In response to your questions about the learned edge weights and their underlying mechanisms (Reviewers bUSc, u5fP, and bZLn), we added edge weight distribution analysis, edge pruning experiments, and domain discriminability analysis. We further examined changes in cross-domain class-pair edge distribution discrepancies before and after reweighting across different relation-view combinations (see the Experiments section and Appendix J).

  • To address your concern about potential projection representation collapse (Reviewers u5fP and bZLn), we analyzed relative class-center distances across domains to examine whether the learned projection representations collapse (see Appendix H).

  • Following your feedback on hyperparameter selection (Reviewer u5fP), we clarified the selection and training protocols, including the predefined number of training epochs and the use of final checkpoints. We also introduced Soft Neighborhood Density (SND) for hyperparameter selection without target-domain ground-truth labels and provided supplementary selection results (see Appendix L).

  • In response to your question about edge-weight symmetry in undirected graphs (Reviewer bZLn), we clarified the bidirectional score averaging mechanism and the use of shared random variables for both directions of each undirected edge to ensure symmetric edge weights. We also specified that Gumbel noise is used only during training, while deterministic edge weights are adopted during inference (see the Method section).

  • Following your suggestion to include an additional baseline (Reviewer bZLn), we added StruRW as a comparative method and reported its experimental results in Tables 2 and 3 of the revised manuscript, providing a more comprehensive comparison with structural reweighting methods (see the Results and Analyses section).

We hope that our responses have sufficiently addressed all of the concerns raised. Should further details or clarifications are needed, please let us know and we would be happy to provide them. Thank you!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.