acceptodds
Under review as a conference paper at ICLR 2027

Class Centroid-Guided Alignment for Multi-View Cross Domain Graphs

Abstract

Multi-view graph domain adaptation (MGDA) addresses label scarcity by transferring knowledge from a labeled source graph to an unlabeled target graph, where nodes are connected through multiple views. Most existing methods align each view as a whole and ignore class-level structure, so that classes can remain misaligned across domains even when view-level distributions match. We empirically reveal that the cross-domain discrepancy in class-level structure is strongly associated with alignment quality. To capture this structure, we use each view's class centroids as anchors and leverage the Cross-Domain Aggregated Class Distance (CD-ACD), which compares source and target inter-class centroid distances within each view and quantifies each view's class-level structural shift. Based on CD-ACD, we propose Class Centroid-Guided Alignment (CCGA), which uses the metric to weight the alignment of each view and class pair during training and thereby reduces class-level structural shifts between domains in each view. Experimental results on various multi-view graph benchmark datasets demonstrate the effectiveness of our method.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph domain adaptation

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.50 over 2 reviews, above 99% of the others.
Reviewer_zJ7wOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work studies unsupervised multi-view graph domain adaptation for cross-domain node classification. The authors propose Cross-Domain Aggregated Class Distance (CD-ACD), a metric that combines cross-domain same-class alignment distances with intra-domain and cross-domain inter-class separation distances to capture class-level structural discrepancies. Building on this metric, the authors develop the Class Centroid-Guided Alignment (CCGA) framework, which dynamically assigns weights to different graph views and integrates graph feature reconstruction, contrastive alignment, source-domain classification, and target-domain entropy minimization. Experiments on the ACM and MAG datasets show that the approach improves classification accuracy and Macro-F1 compared with existing baselines, including SAGA. Ablation studies and hyperparameter sensitivity analyses are also conducted to evaluate the framework.

Critical strengths
  1. This paper tackles a significant challenge in multi-view graph domain adaptation by accounting for class-level structural discrepancies that view-level distribution alignment may fail to capture sufficiently. Incorporating class centroid information into the alignment process provides an intuitive way to enhance cross-domain representation learning.
  2. The proposed method delivers competitive results across multiple transfer tasks on the ACM and MAG datasets, improving both classification accuracy and Macro-F1 relative to existing methods. The experiments also include ablation studies and hyperparameter analyses, offering preliminary evidence for the effectiveness of the proposed framework.
Critical weaknesses
  1. Equation (3) combines cross-domain same-class distances with negatively weighted inter-class separation distances. A lower CD-ACD value may therefore reflect greater class separation as well as closer cross-domain alignment. Both can be useful for adaptation, but this does not by itself establish that CD-ACD reliably measures class-level structural shift. The validation should distinguish these effects more clearly. In addition, although the authors claim to dynamically align individual class pairs, Equations (7)-(9) mainly introduce view-level weighting. How the aggregated class distances lead to explicit alignment of individual class pairs remains unclear. These points leave the interpretation of CD-ACD and its role in class-level alignment insufficiently supported.
  2. Equation (8) (p. 6) is described as a bidirectional InfoNCE contrastive loss, but it does not specify positive pairs or normalize their similarity against competing pairs. As written, the loss can be minimized by mapping all nodes to the same nonzero projected representation, making every cross-domain cosine similarity equal to one. This is a concern about the stated objective, rather than evidence that collapse occurred in the experiments. Since reconstruction and classification operate on the encoder representations, it is unclear what prevents collapse in the projection head. Clarifying whether Equation (8) matches the implementation is necessary to assess the alignment objective.
Questions for authors
  1. Given that Equation (3) combines cross-domain same-class distances with negatively weighted intra-domain and cross-domain inter-class separation distances, so that a lower CD-ACD value may result from increased inter-class separation rather than reduced cross-domain structural discrepancy, can CD-ACD reliably measure the class-level structural shift it is intended to characterize?
  2. Although the authors claim to dynamically align individual class pairs, Equations (7)-(9) primarily introduce view-level weighting. How is class-pair-specific alignment explicitly achieved, and does this weaken the justification for the paper’s central methodological contribution?
  3. Equation (8) is described as a bidirectional InfoNCE contrastive loss, but its formulation does not explicitly distinguish positive from negative sample pairs or include the standard normalization between positive-pair similarity and competing similarities. Does this objective therefore merely encourage cross-domain representational similarity without explicitly preserving class discrimination, raising concerns about the correctness of the proposed alignment objective?
  4. Appendix C states that the experiments were repeated five times, but the main results report only single performance values without standard deviations or other measures of variability. Can the statistical reliability of the reported improvements be assessed?
6 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/4)

Thank you for your insightful comments and constructive suggestions. We address your concerns and questions point by point below.

Response to the Shared Concern in W1 and Q1: Effectiveness of CD-ACD in Measuring Class-Level Structure

Thank you for your insightful comments on whether CD-ACD effectively measures differences in class-level structure.

To further investigate the relationship between CD-ACD and class-level training dynamics, we have added an analysis in Appendix F (EXPERIMENT DETAIL OF EMPIRICAL STUDY) (p. 18, lines 953–960; pp. 18–19, lines 970–974; Figure 3, p. 19, lines 976–1024). For each class, we compute the difference between the source-domain supervised cross-entropy loss and the target-domain prediction entropy, LS−LT\mathcal{L}_S-\mathcal{L}_T, and plot its evolution alongside the minimum CD-ACD score over training epochs. The definitions of these losses are provided in Section 4.3, Eqs. (10)–(11) (p. 6, lines 297–315).

Although the class-specific curves exhibit local fluctuations, they generally decrease during training, accompanied by a corresponding downward trend in the minimum CD-ACD score. This consistency in the observed trends across multiple classes provides supporting evidence that CD-ACD reflects class-level structural information.

In addition, the ablation experiments on CD-ACD components in Appendix G (MODEL ABLATION STUDY) (p. 20, lines 1054–1079; Figure 5, p. 21, lines 1080–1133; Figure 6, p. 22, lines 1134–1174) assess, to some extent, the contributions of the same-class alignment distance and inter-class separation distances to the final model performance.

Table. Ablation results of individual CD-ACD components (%). Each cell reports Accuracy / F1.

MethodACM1→ACM2ACM2→ACM1CN→USUS→CNJP→CNCN→JP
CCGA60.4 / 45.753.8 / 53.361.2 / 43.654.2 / 43.855.5 / 43.859.4 / 47.5
CCGA₁55.8 / 40.247.0 / 47.658.7 / 42.752.9 / 39.854.3 / 42.458.3 / 43.9
CCGA₂34.1 / 29.543.3 / 42.058.2 / 42.252.6 / 39.953.9 / 41.757.8 / 43.3
CCGA₄38.8 / 28.842.5 / 40.257.6 / 42.452.7 / 39.653.2 / 41.558.0 / 43.6
MethodDE→FRFR→DERU→USUS→RUDE→CNCN→DE
CCGA55.8 / 50.057.0 / 50.442.4 / 41.945.9 / 41.947.3 / 43.142.6 / 41.5
CCGA₁54.4 / 45.455.4 / 46.841.6 / 40.544.5 / 40.746.2 / 41.841.7 / 40.5
CCGA₂54.0 / 45.855.1 / 46.641.0 / 39.344.0 / 40.245.7 / 42.241.2 / 39.7
CCGA₄53.9 / 43.054.9 / 46.140.1 / 39.944.1 / 40.245.3 / 40.341.2 / 39.8

CCGA₁ removes the intra-domain class separation distance; CCGA₂ removes the cross-domain cross-class separation distance; CCGA₄ removes the cross-domain same-class alignment distance.

Response to the Shared Concern in W1 and Q2: How CD-ACD Guides Alignment

Thank you for this question and for pointing out the need to further clarify how CD-ACD uses class-level structural information to guide subsequent alignment.

CD-ACD is computed from source-domain and target-domain class centroids. It integrates information about same-class and cross-class pairs within and across domains to select the cross-domain alignment view and generate multi-view consistency weights. In this way, structural information from different classes is aggregated to jointly guide the cross-domain alignment and intra-domain multi-view consistency terms in Eq. (9). The class-centroid construction and score are defined in Section 3 (p. 4, lines 165–198), and the alignment mechanism is described in Section 4.2 (p. 5, lines 261–269; p. 6, lines 270–291).

  • Cross-domain alignment. The model selects the view kk with the lowest CD-ACD score and uses its corresponding weight ωb\omega_b to scale the bidirectional alignment loss. Consequently, cross-domain representation alignment is jointly guided by same-class correspondence, cross-class separation, and intra-domain class separability. See Section 4.2 (p. 6, lines 270–281 and 287–291).

  • Intra-domain multi-view consistency. Within the source and target domains, the model constrains differences in second-order representation statistics between each view and the baseline view. The weights ωv\omega_v determine the contributions of individual views, while 1−ωb1-\omega_b controls the overall weight of this consistency term. Thus, intra-domain consistency is guided not only by statistical differences between views but also by class-level structural information. See Section 4.2 (p. 6, lines 284–293).

AuthorsOpenReview ↗

Official Comment by Authors (2/4)

Response to W2 and Q3: The Bidirectional InfoNCE Contrastive Loss and Class Discriminability

Thank you for your detailed comments on the bidirectional InfoNCE contrastive loss and its class discriminability.

The bidirectional InfoNCE contrastive loss presented in the paper does not explicitly construct same-class positive pairs, cross-class negative pairs, or class-dependent masks. The original loss is given in Section 4.2, Eq. (8) (p. 6, lines 275–281). To investigate class discriminability during actual training, we analyze distances between different classes in the prediction space and provide the full explanation in Appendix K (CLASS DISCRIMINABILITY ANALYSIS) (p. 27–28, lines 1406–1510; Tables 4–7). Selected results are shown in Table 1. Different classes retain distinguishable distances in the prediction space, indicating that the final predictions have not degenerated into identical outputs across all classes and retain class discriminability.

Table 1. Relative class representation distances in the prediction space on CN→DE under view 1 with the original loss.

S/TC0C1C2C3C4C5C6C7C8C9
C01.005.125.802.484.554.573.386.011.571.59
C14.551.002.135.172.572.385.632.284.704.59
C24.962.001.005.512.412.215.941.255.064.98
C32.765.325.891.004.954.981.746.003.053.33
C42.891.551.633.371.000.793.751.902.962.91
C53.802.032.204.411.401.004.912.553.893.82
C62.595.115.580.904.694.701.005.702.813.13
C74.141.731.044.592.232.074.931.004.264.18
C81.247.028.023.676.306.294.958.411.001.24
C90.714.164.742.163.713.722.884.971.121.00
AuthorsOpenReview ↗

Official Comment by Authors (3/4)

Response to W2 and Q3: The Bidirectional InfoNCE Contrastive Loss and Class Discriminability

To further investigate the effect of explicitly constructing same-class positive pairs and cross-class negative pairs with class-dependent masks, we design a bidirectional InfoNCE contrastive loss incorporating class encoding.

l(ZkS,ZkT)=−1∣IS∣∑i∈ISlog⁡(∑j=1NTMij+exp⁡(sijS,T/τ)∑j=1NT(Mij++Mij−)exp⁡(sijS,T/τ))l(Z_k^S,Z_k^T)=-\frac{1}{|I_S|}\sum_{i\in I_S}\log\left(\frac{\sum_{j=1}^{N_T}M_{ij}^{+}\exp(s_{ij}^{S,T}/\tau)}{\sum_{j=1}^{N_T}(M_{ij}^{+}+M_{ij}^{-})\exp(s_{ij}^{S,T}/\tau)}\right)

We denote this variant with explicit class-aware positive and negative pairs as CCGA-CP and compare its class discriminability and classification performance with CCGA, which uses the original loss presented in the paper. The full experimental description has been added to Appendix L (COMPARISON OF BIDIRECTIONAL ALIGNMENT LOSSES) (p. 29–31, lines 1514–1634; Table 8), with selected results shown in Tables 2 and 3. These correspond to Table 10 (p. 30, lines 1589–1602) and Table 8 (p. 29, lines 1550–1565), respectively, in Appendix L.

The results show that explicit positive and negative sample constraints improve the measured class discriminability but do not yield corresponding overall improvements in Accuracy or F1. This indicates that stronger explicit class contrastive constraints do not necessarily translate into better target-domain classification performance under the current experimental settings, and may instead introduce additional computational overhead.

Table 2. Relative class representation distances in the prediction space on CN→DE under view 1 with the new loss.

S/TC0C1C2C3C4C5C6C7C8C9
C01.0014.3112.3710.1910.1010.3613.9712.721.151.24
C18.341.005.946.856.976.947.685.928.468.38
C2144.67113.771.00189.9640.3037.10234.064.75145.87143.17
C38.7610.5012.921.0012.1812.332.7213.118.789.01
C425.8829.717.9039.661.000.2950.068.8326.0125.48
C576.9987.8723.04117.703.311.00148.5225.8277.3875.80
C651.3649.9266.7314.4564.9565.551.0067.4351.5952.48
C758.4643.712.1475.3717.8616.6192.491.0058.9657.88
C81.2430.4824.4822.7119.1619.7631.2525.261.001.40
C90.7222.6218.9216.8515.1915.6223.1019.501.191.00

Table 3. Experimental results of CCGA with different loss functions (%).

MethodMetricA1→A2A2→A1CN→USUS→CNJP→CNCN→JP
CCGAAcc60.4±0.3553.8±2.5761.2±1.0854.2±0.5455.5±0.3359.4±0.22
CCGAF145.7±0.3953.3±2.5643.6±0.5043.8±0.3043.8±0.6547.5±0.30
CCGA-CPAcc60.5±0.4053.6±6.8761.1±0.6553.9±0.2255.4±0.4459.4±0.13
CCGA-CPF145.5±0.3354.0±6.8743.9±1.2444.1±0.3743.9±0.5147.7±0.29
MethodMetricDE→FRFR→DERU→USUS→RUDE→CNCN→DE
CCGAAcc55.8±0.2457.0±0.2642.4±0.4645.9±0.2447.3±0.7342.6±0.29
CCGAF150.0±0.1250.4±0.6841.9±0.6341.9±0.2543.1±1.0441.5±0.21
CCGA-CPAcc55.6±0.3157.1±0.1142.3±1.2445.9±0.1947.1±0.2942.7±0.24
CCGA-CPF150.1±0.1750.5±0.1642.0±1.1141.8±0.2143.2±0.3641.6±0.28
Reviewer_1KzKOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper studies Multi-view Graph Domain Adaptation (MVGDA) for the node classification task. Since existing works ignore the class-level structure and only align the view-level distributions, a metric quantifying each view's class-level structural shift, Cross-Domain Aggregated Class Distance (CD-ACD) , is designed to bridge this gap. Leveraging CD-ACD, the CCGA algorithm is proposed to mitigate the structural discrepancies of multi-view graphs at both class-level and view-level. Extensive experiments on various baselines demonstrate the superior performance of CCGA.

Critical strengths
  1. This paper is logically coherent and easy to follow. It analyses a clear research gap in MVGDA, namely the ignorance of class-level structural distribution, and then empirically shows that CD-ACD has better performance than SAD in mitigating distribution shift. In addition, the figures are clear and informative, providing a good visualization.

  2. The method designs the class centroid for each class and each view, which is applied to measure the structural distribution shift at both class-level and view-level, thus addressing the limitation of existing works. Experimental results on multiple baselines and real-world cross-domain datasets with accuracy and macro-F1 metrics validate that CCGA is highly effective in MVGDA node classification tasks.

Critical weaknesses
  1. This paper lacks ablation study to assess the contribution of cross-domain same-class alignment.
  2. The range of the proposed metric CD-ACD_v should be clarified. It is unclear whether its sign would affect subsequent stages.
Questions for authors
  1. Can the authors provide an additional ablation study to show the performance removing the cross-domain same-class alignment distance? Or replace it with another conventional metric to measure cross-domain same-class distance.
  2. Is it possible for CD-ACD_v to be negative? What is the range of this term? And if it may be positive or negative at the same time, will the following stage based on it be affected?
  3. Why does Eq.(7) use negative lambda instead of lambda itself?
  4. Please clarify what are the differences between the MVGDA problem and heterogenous GDA problem (i.e., GDA designed for heterogenous graphs). Whether the heterogenous GDA methods can be directly applied to MVGDA?
5 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/3)

Thank you for your insightful comments and constructive suggestions. We address your concerns and questions point by point below.

Response to W1 and Q1: Ablation Study on the Cross-Domain Same-Class Alignment Distance

Thank you for this suggestion. In Appendix G (MODEL ABLATION STUDY) (p. 20, lines 1075–1079; Figure 6, p. 22, lines 1134–1174), we have added an ablation variant, denoted as CCGA₄, which removes the cross-domain same-class alignment distance from CD-ACD, and compared it with the full CCGA model.

The following tables report Accuracy and F1 across all twelve transfer tasks. All values are expressed as percentages (%).

MethodMetricACM1→ACM2ACM2→ACM1CN→USUS→CNJP→CNCN→JP
CCGAAccuracy60.453.861.254.255.559.4
CCGAF145.755.343.643.843.847.5
CCGA₄Accuracy38.842.557.652.753.258.0
CCGA₄F128.840.242.439.641.543.6
MethodMetricDE→FRFR→DERU→USUS→RUDE→CNCN→DE
CCGAAccuracy55.857.042.445.947.342.6
CCGAF150.050.441.941.943.141.5
CCGA₄Accuracy53.954.940.144.145.341.2
CCGA₄F143.046.139.940.240.339.8

The results show that removing this component leads to a decline in both Accuracy and F1 across all twelve transfer tasks, indicating that the cross-domain same-class alignment distance provides useful information about corresponding classes for CD-ACD.

AuthorsOpenReview ↗

Official Comment by Authors (2/3)

Response to W2 and Q2: The Range of CD-ACD and Its Effect on Subsequent Stages

Thank you for pointing out the need to clarify the range of CD-ACD and the effect of its sign on subsequent stages.

  • Empirical observations and mathematical properties. As shown in Appendix F (EXPERIMENT DETAIL OF EMPIRICAL STUDY) (pp. 18–19, lines 970–974; Figure 3, p. 19, lines 976–1024; Figure 4, p. 20, lines 1026–1047), CD-ACD remains positive throughout training in the examined experiments with the full model. However, the positive values observed experimentally do not imply that Eq. (3) mathematically guarantees non-negativity. CD-ACD consists of a non-negative same-class distance term minus weighted inter-class separation terms and can therefore take negative values. After removing the same-class distance term, the remaining score is negative in our experiments.

  • Range. The paper adopts multi-view mean normalization. For each non-negative distance component with a nonzero mean across views, every normalized value lies in [0,V][0,V], where VV is the number of views. Therefore, provided that γ1,γ2≥0\gamma_1,\gamma_2\geq 0 and all normalization denominators for the distance components are nonzero, CD-ACD satisfies the following bound (Section 3, p. 4, lines 188–198, including Eq. (3)):

    \leq \mathrm{CD\text{-}ACD} \leq V.$$ However, negative CD-ACD scores are undesirable in MGDA, and thus we retain the cross-domain same-class alignment distance term in the formulation of CD-ACD.
  • Effect on subsequent computations. Negative CD-ACD scores do not affect the Softmax-based view weight assignment in Eq. (7), since the weights are determined by the relative CD-ACD scores across views. The weighting rule and its interpretation are given in Section 4.2 (p. 5, lines 261–269; p. 6, lines 270–274).

AuthorsOpenReview ↗

Official Comment by Authors (3/3)

Response to Q3: The Choice of Negative Rather Than Positive λ

Thank you for this question. Eq. (7) uses −λ-\lambda in the exponent, with λ>0\lambda>0, so that views with lower CD-ACD scores receive higher weights ω\omega. See Section 4.2 (p. 5, lines 261–269; p. 6, lines 270–274).

The rationale behind our design is that a lower CD-ACD score indicates a smaller cross-domain class-level structural shift and relatively better cross-domain alignment quality for that view. Such a view should therefore receive a higher weight during subsequent multi-view alignment. With −λ-\lambda, a higher CD-ACD score produces a smaller exponential term and consequently a lower normalized weight ω\omega, consistent with this design objective. Using +λ+\lambda would instead assign higher weights to views with higher CD-ACD scores, contrary to the intended behavior.

Response to Q4: Differences Between MVGDA and Heterogeneous GDA

Thank you for this question. Both MVGDA and heterogeneous GDA leverage graph structural information to facilitate cross-domain knowledge transfer, but they differ in their study objects and alignment focuses.

  • Study objects and problem settings. In this paper, MVGDA takes multiple graph views over the same set of nodes to be classified as input, where different views describe different structural relations among these nodes. It requires jointly considering the cross-domain distribution shifts across multiple views. In contrast, heterogeneous GDA typically takes a single heterogeneous graph containing multiple node types or edge types as its modeling object, focusing on cross-domain differences in type-specific attributes, relations, and semantic information between the source and target domains.

  • Structural modeling and alignment focus. MVGDA performs structural modeling and cross-domain alignment across multiple views, considering both the complementary information provided by each view and their different contributions to knowledge transfer[1]. Our method further utilizes class-level structural information from each view to guide the selection of cross-domain alignment views and dynamic weight assignment. Heterogeneous GDA, on the other hand, primarily models the semantics of different node and relation types within a single heterogeneous graph and performs cross-domain alignment based on these representations. Multiple relation types are not equivalent to multiple independent graph views[2].

  • Commonalities. Both settings leverage node attributes and structural relations to mitigate distribution shifts between the source and target domains, thereby improving node classification performance in the target domain. Therefore, they share common ground in utilizing relational information and learning cross-domain representations. However, MVGDA places greater emphasis on the collaboration among multiple views and their different transfer contributions, whereas heterogeneous GDA focuses more on type and relation semantics within a single graph.

References

[1] Ruiyi Fang, Bingheng Li, Jingyu Zhao, Ruizhi Pu, Qiuhao Zeng, Gezheng Xu, Charles Ling, and Boyu Wang. Homophily Enhanced Graph Domain Adaptation. In Proceedings of the 42nd International Conference on Machine Learning, PMLR 267, pp. 16006–16028, 2025.

[2] Boshen Shi, Yongqing Wang, Fangda Guo, Jiangli Shao, Huawei Shen, and Xueqi Cheng. Improving Graph Domain Adaptation with Network Hierarchy. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pp. 2249–2258, 2023.

Comments on the paper

5 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/3)

Thank you for your insightful comments and constructive suggestions. We address your concerns and questions point by point below.

Response to W1 and Q1: The Bidirectional InfoNCE Contrastive Loss and class-aware alignment

Thank you for your detailed comments on the alignment loss, the role of class information, and the gradient propagation paths.

CD-ACD is computed from source-domain and target-domain class centroids. It integrates information about same-class and cross-class pairs within and across domains to select the cross-domain alignment view and generate multi-view consistency weights. In this way, structural information from different classes is aggregated to jointly guide the cross-domain alignment and intra-domain multi-view consistency terms in Eq. (9). The class-centroid construction and score are defined in Section 3 (p. 4, lines 165–198), and the alignment mechanism is described in Section 4.2 (p. 5, lines 261–269; p. 6, lines 270–291).

  • Forward propagation. The model generates node representations and target-domain prediction probabilities, then uses source-domain labels and target-domain confidence weights to compute class centroids, CD-ACD scores, and Softmax view weights. These quantities are subsequently used to select the cross-domain alignment view and construct the joint alignment loss.

  • Backward propagation. Target-domain prediction probabilities are treated as fixed confidence weights. Gradients propagate directly through the representation discrepancy terms and also through the path “Softmax weights → CD-ACD → class centroids → encoder representations.”

Specifically, when confidence filtering rejects all target nodes, such that the sum of the filtered target weights does not exceed the numerical threshold ε\varepsilon, the implementation falls back to soft confidence weighting for the current batch. This fallback removes the hard confidence threshold while retaining “class prediction probability × prediction confidence raised to the power γpow\gamma_{\mathrm{pow}}” as the weight. We have added details of the target-domain centroids and the associated configurations in Appendix M (ADDITIONAL EXPERIMENTAL DETAILS) (p. 31, lines 1655–1671).

The bidirectional InfoNCE contrastive loss presented in the paper does not explicitly construct same-class positive pairs, cross-class negative pairs, or class-dependent masks. The original loss is given in Section 4.2, Eq. (8) (p. 6, lines 275–281). To investigate the discriminability of class representations during actual training, we analyze relative distances between different classes in the projection space and provide the full explanation in Appendix K (CLASS DISCRIMINABILITY ANALYSIS) (p. 27–28, lines 1406–1510; Tables 4–7). Selected results are shown in Table 1. Different classes retain distinguishable distances in the projection space, indicating that the final predictions have not degenerated into identical outputs across all classes and retain class discriminability. The corresponding result is Table 5 in Appendix K (p. 28, lines 1459–1474).

Table 1. Relative class representation distances in the projection space on CN→DE under view 1 with the original loss.

S/TC0C1C2C3C4C5C6C7C8C9
C01.005.125.802.484.554.573.386.011.571.59
C14.551.002.135.172.572.385.632.284.704.59
C24.962.001.005.512.412.215.941.255.064.98
C32.765.325.891.004.954.981.746.003.053.33
C42.891.551.633.371.000.793.751.902.962.91
C53.802.032.204.411.401.004.912.553.893.82
C62.595.115.580.904.694.701.005.702.813.13
C74.141.731.044.592.232.074.931.004.264.18
C81.247.028.023.676.306.294.958.411.001.24
C90.714.164.742.163.713.722.884.971.121.00
AuthorsOpenReview ↗

Official Comment by Authors (2/3)

Response to W1 and Q1: The Bidirectional InfoNCE Contrastive Loss and class-aware alignment

To further investigate the effect of explicitly constructing same-class positive pairs and cross-class negative pairs with class-dependent masks, we design a bidirectional InfoNCE contrastive loss incorporating class encoding.

l(ZkS,ZkT)=−1∣IS∣∑i∈ISlog⁡(∑j=1NTMij+exp⁡(sijS,T/τ)∑j=1NT(Mij++Mij−)exp⁡(sijS,T/τ))l(Z_k^S,Z_k^T)=-\frac{1}{|I_S|}\sum_{i\in I_S}\log\left(\frac{\sum_{j=1}^{N_T}M_{ij}^{+}\exp(s_{ij}^{S,T}/\tau)}{\sum_{j=1}^{N_T}(M_{ij}^{+}+M_{ij}^{-})\exp(s_{ij}^{S,T}/\tau)}\right)

We denote this variant with explicit class-aware positive and negative pairs as CCGA-CP and compare its class discriminability and classification performance with CCGA, which uses the original loss presented in the paper. The full experimental description has been added to Appendix L (COMPARISON OF BIDIRECTIONAL ALIGNMENT LOSSES) (p. 29–31, lines 1514–1634; Table 8), with selected results shown in Tables 2 and 3. These correspond to Table 10 (p. 30, lines 1589–1602) and Table 8 (p. 29, lines 1550–1565), respectively, in Appendix L.

The results show that explicit positive and negative sample constraints improve the measured class discriminability but do not yield corresponding overall improvements in Accuracy or F1. This indicates that stronger explicit class contrastive constraints do not necessarily translate into better target-domain classification performance under the current experimental settings, and may instead introduce additional computational overhead.

Table 2. Relative class representation distances in the projection space on CN→DE under view 1 with the new loss.

S/TC0C1C2C3C4C5C6C7C8C9
C01.0014.3112.3710.1910.1010.3613.9712.721.151.24
C18.341.005.946.856.976.947.685.928.468.38
C2144.67113.771.00189.9640.3037.10234.064.75145.87143.17
C38.7610.5012.921.0012.1812.332.7213.118.789.01
C425.8829.717.9039.661.000.2950.068.8326.0125.48
C576.9987.8723.04117.703.311.00148.5225.8277.3875.80
C651.3649.9266.7314.4564.9565.551.0067.4351.5952.48
C758.4643.712.1475.3717.8616.6192.491.0058.9657.88
C81.2430.4824.4822.7119.1619.7631.2525.261.001.40
C90.7222.6218.9216.8515.1915.6223.1019.501.191.00

Table 3. Experimental results of CCGA with different loss functions (%).

MethodMetricA1→A2A2→A1CN→USUS→CNJP→CNCN→JP
CCGAAcc60.4±0.3553.8±2.5761.2±1.0854.2±0.5455.5±0.3359.4±0.22
CCGAF145.7±0.3953.3±2.5643.6±0.5043.8±0.3043.8±0.6547.5±0.30
CCGA-CPAcc60.5±0.4053.6±6.8761.1±0.6553.9±0.2255.4±0.4459.4±0.13
CCGA-CPF145.5±0.3354.0±6.8743.9±1.2444.1±0.3743.9±0.5147.7±0.29
MethodMetricDE→FRFR→DERU→USUS→RUDE→CNCN→DE
CCGAAcc55.8±0.2457.0±0.2642.4±0.4645.9±0.2447.3±0.7342.6±0.29
CCGAF150.0±0.1250.4±0.6841.9±0.6341.9±0.2543.1±1.0441.5±0.21
CCGA-CPAcc55.6±0.3157.1±0.1142.3±1.2445.9±0.1947.1±0.2942.7±0.24
CCGA-CPF150.1±0.1750.5±0.1642.0±1.1141.8±0.2143.2±0.3641.6±0.28
AuthorsOpenReview ↗

Official Comment by Authors (3/3)

Response to W2 and Q2: Effectiveness Analysis of CD-ACD

Thank you for suggesting matched-backbone control experiments to further assess the contribution of CD-ACD and for pointing out the need to clarify the definition of LT\mathcal{L}_T in Appendix F.

  • Matched-backbone control experiments. To assess the contribution of CD-ACD beyond the remaining components, we retain the same model backbone and loss components and introduce three control variants: (1) fixed view, which fixes the cross-domain alignment view to view 0; (2) fixed weight, which replaces the dynamic view weights with uniform weights; and (3) fixed view & weight, which fixes both the alignment view and the view weights. We compare these three variants with the full model (full CCGA) and provide the complete experimental description in Appendix J (EFFECTIVENESS ANALYSIS OF CD-ACD) (p. 25–26, lines 1326–1403; Figure 8). Selected results are presented here.

Table 4. Matched-backbone control results (%).

MethodMetricACM1→ACM2ACM2→ACM1CN→USUS→CN
full CCGAAcc60.453.861.254.2
full CCGAF145.755.343.643.8
fixed viewAcc47.445.558.452.7
fixed viewF136.245.741.743.3
fixed weightAcc45.744.457.952.6
fixed weightF135.243.241.942.6
fixed view & weightAcc43.441.755.152.1
fixed view & weightF133.840.140.941.2

The results show that, on the ACM1 and MAG datasets, full CCGA achieves clear improvements in both Accuracy and F1 over the variants with a fixed view, fixed weights, or both. Fixing either the view or the weights reduces performance, while fixing both yields the lowest performance. These results indicate that CD-ACD-guided view selection and dynamic weight assignment both contribute to the final performance. The gains arise from effectively guiding the existing components, rather than solely from the remaining model components. In addition, the ablation experiments on CD-ACD components in Appendix G (MODEL ABLATION STUDY) (p. 20–22, lines 1054–1174; Figure 5) assess, to some extent, the contributions of the same-class alignment distance and inter-class separation distances to the final model performance.

  • Class-level training dynamics and the definition of LT\mathcal{L}_T. To further investigate the role of CD-ACD during class-level alignment training, we have added an analysis in Appendix F (EXPERIMENT DETAIL OF EMPIRICAL STUDY) (p. 18–19, lines 953–1024; Figure 3). For each class, we compute the difference between the source-domain supervised cross-entropy loss and the target-domain prediction entropy, LS−LT\mathcal{L}_S-\mathcal{L}_T, and plot its evolution alongside the minimum CD-ACD score over training epochs. Although the class-specific curves exhibit local fluctuations, they generally decrease during training, accompanied by a corresponding downward trend in the minimum CD-ACD score. This consistency in the observed trends across multiple classes provides supporting evidence that CD-ACD reflects training dynamics related to class-level structure. We also thank the reviewer for specifically pointing out the issue regarding the definition of LT\mathcal{L}_T. We have explicitly defined LT\mathcal{L}_T as the target-domain prediction entropy loss in the manuscript and corrected typographical errors involving notations such as LS\mathcal{L}^S and LT\mathcal{L}^T throughout the paper. The source supervised loss and target prediction entropy loss are defined in Section 4.3, Eqs. (10)–(11) (p. 6, lines 297–322), and the latter is restated in Appendix F (p. 18, lines 953–960).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.