- 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 . APPNP propagation is then performed on the reweighted graphs, followed by a shared MLP encoder.
Critical strengths
-
The research question is interesting and valuable.
-
The paper is well-written, organized, and easy to follow
-
To the best of my knowledge, this is a novel approach.
-
Using the downstream alignment objective to guide message-passing edge weights is more adaptive than propagating over a fixed graph.
-
Bounding the learned edge weights within 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
-
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 , where is any nonzero vector. All cosine similarities then equal one, attaining the global optimum of the objective. Because classification is performed using , whereas alignment is performed through separate projection heads , the source classification loss does not necessarily prevent the projection heads from collapsing.
-
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.
-
The edge score is computed as , which is generally not invariant to swapping and . Therefore, and may receive different weights. This is inconsistent with the undirected-graph assumption and symmetric normalization unless the weights are explicitly symmetrized.
Questions for authors
-
How does the method avoid the collapsed solution of Eq. (9), in which every projected source and target representation is identical?
-
Can the authors report the learned edge-weight distributions and quantify their relationship with edge transferability?
-
Do the class-pair edge-distribution discrepancies shown in Figure 1 decrease after edge reweighting?
-
Is the constraint explicitly enforced for undirected edges?
-
Why is the closest structural reweighting graph domain adaptation method not included as a baseline?
-
Is Gumbel noise still sampled during inference? If so, how is prediction stability ensured?
11 comments
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 -th view of domain , let denote the set of nodes belonging to class , and denote the corresponding number of nodes. The class center is defined as:
Here, denotes the projected representation of node in the -th view. For source class and target class , 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, and denote the projected class centers of source class and target class in the -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/T | C0 | C1 | C2 | C3 | C4 | C5 | C6 | C7 | C8 | C9 |
|---|---|---|---|---|---|---|---|---|---|---|
| C0 | 1.00 | 7.58 | 5.36 | 8.08 | 5.70 | 5.14 | 9.81 | 6.08 | 2.89 | 2.40 |
| C1 | 3.40 | 1.00 | 2.80 | 5.13 | 3.11 | 3.82 | 5.90 | 2.98 | 3.96 | 3.68 |
| C2 | 3.78 | 4.56 | 1.00 | 6.15 | 2.88 | 2.39 | 7.57 | 1.80 | 4.35 | 4.31 |
| C3 | 11.20 | 15.70 | 12.05 | 1.00 | 11.39 | 10.39 | 15.52 | 11.67 | 9.90 | 11.21 |
| C4 | 2.08 | 3.74 | 2.40 | 4.07 | 1.00 | 0.75 | 5.86 | 2.32 | 3.28 | 3.47 |
| C5 | 4.06 | 4.49 | 2.84 | 6.36 | 1.51 | 1.00 | 7.96 | 2.73 | 4.51 | 4.76 |
| C6 | 6.00 | 8.98 | 7.18 | 2.27 | 6.05 | 6.09 | 1.00 | 6.95 | 6.34 | 6.59 |
| C7 | 3.10 | 3.36 | 1.97 | 4.30 | 2.15 | 1.94 | 5.34 | 1.00 | 3.35 | 3.51 |
| C8 | 2.01 | 6.67 | 4.88 | 5.12 | 4.72 | 4.32 | 6.51 | 5.18 | 1.00 | 1.59 |
| C9 | 0.47 | 4.38 | 3.10 | 4.66 | 3.24 | 2.90 | 5.69 | 3.46 | 1.43 | 1.00 |
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.
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 -th view:
Here, denotes the projected representation of node in the -th view of domain , contains source nodes of class , and contains target nodes with high-confidence pseudo-labels of class . denotes the corresponding number of nodes. We then employ symmetric cross-entropy to encourage cross-domain same-class matching and different-class discrimination:
Here, 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:
Here, denotes the -th component of node 's projected representation in the -th view, is its mean, is the number of nodes used for computation in the -th view of domain , is the projection dimension, is the standard deviation threshold, and 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/T | C0 | C1 | C2 | C3 | C4 | C5 | C6 | C7 | C8 | C9 |
|---|---|---|---|---|---|---|---|---|---|---|
| C0 | 1.00 | 41.93 | 24.19 | 23.41 | 22.59 | 22.67 | 29.33 | 22.61 | 3.65 | 1.80 |
| C1 | 30.81 | 1.00 | 14.57 | 15.42 | 15.98 | 15.91 | 11.54 | 15.70 | 29.08 | 30.31 |
| C2 | 74.08 | 59.16 | 1.00 | 13.18 | 9.12 | 8.95 | 21.23 | 6.19 | 66.18 | 71.80 |
| C3 | 32.30 | 27.02 | 5.80 | 1.00 | 5.87 | 5.79 | 7.45 | 6.20 | 28.73 | 31.32 |
| C4 | 13.74 | 11.34 | 1.41 | 2.29 | 1.00 | 0.94 | 3.97 | 1.80 | 12.25 | 13.31 |
| C5 | 22.86 | 19.95 | 2.44 | 3.82 | 1.12 | 1.00 | 7.27 | 2.72 | 20.32 | 22.13 |
| C6 | 43.17 | 23.50 | 10.00 | 8.47 | 11.78 | 11.64 | 1.00 | 11.92 | 39.23 | 42.07 |
| C7 | 21.47 | 18.38 | 0.93 | 3.94 | 2.64 | 2.61 | 7.00 | 1.00 | 19.11 | 20.79 |
| C8 | 4.17 | 46.25 | 24.37 | 23.35 | 22.39 | 22.48 | 30.61 | 22.46 | 1.00 | 3.26 |
| C9 | 0.27 | 48.83 | 27.64 | 26.72 | 25.73 | 25.82 | 33.78 | 25.75 | 3.18 | 1.00 |
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.
| Task | ACC (CARE) | ACC (CARE-1) | Macro-F1 (CARE) | Macro-F1 (CARE-1) |
|---|---|---|---|---|
| CN-US | 0.593 | 0.597 | 0.420 | 0.426 |
| JP-CN | 0.532 | 0.531 | 0.402 | 0.400 |
| CN-JP | 0.577 | 0.582 | 0.446 | 0.451 |
| DE-FR | 0.543 | 0.548 | 0.441 | 0.444 |
| RU-US | 0.389 | 0.389 | 0.387 | 0.386 |
| ACM1-ACM2 | 0.538 | 0.543 | 0.432 | 0.437 |
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.
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.