acceptodds
Under review as a conference paper at ICLR 2027

DEAR: Decision Boundary-Aware Open-set Graph Domain Adaptation

Abstract

Graph domain adaptation, a subfield of graph transfer learning, aims to transfer knowledge from a label-rich source domain to a related but label-scarce target domain. In the Open-Set Graph Domain Adaptation (OSGDA) setting, the target domain contains novel classes that are absent in the source domain, which introduces significant difficulty by interfering with the recognition of known classes. How to effectively transfer domain-invariant knowledge while distinguishing unknown classes remains a major challenge. Existing OSGDA approaches primarily focus on separating unknown classes in the target domain, often overlooking the limitations imposed by the robustness of representation learning in the source domain itself. To address this, we propose a novel decision boundary-aware method called DEAR. Specifically, we introduce a virtual class to compress the decision space of known classes, thereby reserving space in the decision boundary for unknown classes in the target domain. Then, within the target domain, we exploit the virtual class to mine intra-class relationships and accurately identify the decision boundary between known and unknown classes. Finally, we perform cross-domain contrastive learning based on feature and structure perturbations for the known-class nodes to further align transferable knowledge. Extensive experiments on multiple benchmark datasets demonstrate that achieves state-of-the-art performance.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Networks, Graph Domain Adaptation, Open-set Learning

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 4.00 over 2 reviews, above 99% of the others.
Reviewer_TLgwOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a decision boundary aware framework for open set graph domain adaptation. It employs virtual class learning to enhance source domain discriminability, followed by mining class relation to distinguish known and unknown target nodes. Moreover, it employs contrastive learning with perturbations to facilitate cross domain knowledge transfer. Both theoretical analyses and experiments demonstrate the effectiveness of the framework.

Critical strengths
  1. The paper addresses the practically relevant OSGDA problem, where the target domain contains novel classes absent from the source domain.
  2. The method improves knowledge transfer by optimizing source domain decision boundaries, rather than focusing solely on unknown class separation in the target domain.
  3. Virtual class learning is well suited to OSGDA, as it enhances known class discriminability and supports the separation of known and unknown classes.
  4. The paper provides theoretical analyses to support the proposed virtual class learning mechanism, offering theoretical insights.
Critical weaknesses
  1. This paper adopts overall accuracy and HS as the evaluation metrics for OSGDA, which focuses on the overall performance on known and unknown classes. It is unclear how the proposed method performs on known class classification and unknown class detection, respectively, in OS* (known accuracy) and AUC, compared with OSGDA baselines.
  2. The unknown class identification mechanism relies on the assumption that unknown target nodes generally exhibit larger KL divergence from source domain class prototypes. However, this assumption may not always hold under substantial domain shifts or when unknown classes are semantically similar to known classes. In such cases, the divergence distributions of known and unknown nodes may overlap significantly.
  3. For threshold selection, Eq. (12) employs a fixed threshold of ϵ=0.5\epsilon=0.5 to select high confidence known nodes. However, different datasets may have different representation distributions. The paper does not justify this threshold or investigate its sensitivity, leaving its general applicability unclear. In addition, in Eq. (9), it is unclear how to “adaptively determine the threshold according to the underlying data distribution”.
Questions for authors
  1. How the proposed method performs on known class classification and unknown class detection, respectively, compared with the OSGDA baselines? Please provide comparison between the proposed method and OSGDA baselines in OS* (known accuracy) and AUC on some representative tasks.
  2. How do improved source domain decision boundaries facilitate the recognition of unseen target classes and under what conditions is this mechanism effective?
  3. This paper uses a fixed η\eta-quantile of a fitted beta distribution to identify the unknown nodes, how does the fixed quantile parameter η\eta affect unknown-node identification when the proportion of unknown nodes varies across target domains?
  4. In Eq. (9), how to adaptively determine the threshold according to the underlying data distribution?
  5. Why do the authors adopt a single beta distribution to model the KL divergence values of both known and unknown target nodes? Can a single beta distribution adequately model the divergence values of both known and unknown nodes, compared with alternatives such as gaussian mixture models?
  6. For the edge down-weighting strategy, if known nodes are incorrectly identified as unknown, their associated edges will also be down-weighted, potentially suppressing useful neighborhood information. How does this affect subsequent message passing, and can the proposed method mitigate this issue?
5 comments

We sincerely thank you for your time and effort in reviewing our manuscript and for providing valuable feedback. We carefully address your comments and concerns in the following responses.

Q1: Experimental Results regarding OS* and AUC.

R1: Thank you for your valuable comment! To further evaluate the performance of our DEAR and the OSGDA baselines (SDA[1], UAGA[2], ETA[3], and GraphRTA[4]), we conduct additional experiments on six transfer tasks, treating classes 0 and 1 as unknown classes, and each experiment is repeated five times, and we report the mean and standard deviation of known-class classification accuracy (OS∗OS^*) and AUC. The results are presented below (with the best score highlighted in bold):

OS∗OS^*SDAUAGAETAGraphRTADEAR
A2C55.32 ±\pm 0.8474.01 ±\pm 4.3370.86 ±\pm 2.6879.17 ±\pm 2.6173.78 ±\pm 4.29
A2D46.66 ±\pm 1.5165.45 ±\pm 3.7771.43 ±\pm 1.2670.98 ±\pm 3.1375.44 ±\pm 4.06
C2D53.31 ±\pm 0.9173.56 ±\pm 3.7365.86 ±\pm 3.5873.08 ±\pm 2.1669.36 ±\pm 3.41
C2A55.34 ±\pm 0.7477.15 ±\pm 6.7572.53 ±\pm 2.3274.54 ±\pm 4.6975.76 ±\pm 3.90
D2A49.64 ±\pm 1.6056.01 ±\pm 3.3257.49 ±\pm 1.8461.04 ±\pm 1.5462.20 ±\pm 2.07
D2C49.98 ±\pm 1.2069.43 ±\pm 2.7968.84 ±\pm 3.4671.15 ±\pm 1.7472.77 ±\pm 2.93
AUCSDAUAGAETAGraphRTADEAR
A2C79.31 ±\pm 0.9080.63 ±\pm 2.6481.23 ±\pm 2.3678.63 ±\pm 0.7186.31 ±\pm 1.34
A2D77.09 ±\pm 2.5682.44 ±\pm 3.7283.54 ±\pm 2.2874.08 ±\pm 0.9686.43 ±\pm 1.98
C2D82.40 ±\pm 0.2784.87 ±\pm 2.9179.48 ±\pm 3.0877.17 ±\pm 1.4686.93 ±\pm 1.53
C2A71.60 ±\pm 2.1572.83 ±\pm 1.1574.54 ±\pm 1.9865.44 ±\pm 0.8777.13 ±\pm 2.90
D2A59.34 ±\pm 2.1162.83 ±\pm 2.3662.04 ±\pm 2.2154.32 ±\pm 0.7867.66 ±\pm 0.69
D2C71.30 ±\pm 1.3979.56 ±\pm 2.3472.29 ±\pm 3.3567.83 ±\pm 1.8178.74 ±\pm 1.10

As shown above, existing OSGDA methods generally achieve competitive performance in known-class classification, while our DEAR maintains comparable OS∗OS^* performance and demonstrates particularly strong results in AUC, achieving the best performance on most transfer tasks. This indicates that DEAR is more effective at distinguishing known-class nodes from unknown-class nodes across different decision thresholds. We attribute this improvement to the virtual-class guided boundary learning, which encourages more discriminative representations of known classes, together with the KL divergence based unknown-class identification mechanism, which facilitates the separation of known and unknown target nodes. These findings further demonstrate the effectiveness of our method in OSGDA tasks.

We hope these additional experiments adequately address your concerns.

[1] Open-set graph domain adaptation via separate domain alignment, 2024 AAAI.

[2] Open-Set Cross-Network Node Classification via Unknown-Excluded Adversarial Graph Domain Alignment, 2025 AAAI.

[3] ETA: Dual Evidence-Aware Uncertainty Learning for Open-Set Graph Domain Adaptation, 2025.

[4] Towards Unsupervised Open-Set Graph Domain Adaptation via Dual Reprogramming, 2025 NeurIPS.

Q2: Discussion about the Effectiveness of Improved Decision Boundaries.

R2: Thank you for your insightful question! The improved source-domain decision boundaries facilitate unknown-class recognition by enhancing the discriminability of known-class representations. As analyzed in Theorem 1, virtual-class learning introduces additional feature-alignment supervision that facilitates classification margin optimization, encouraging more discriminative known-class representations and source-class prediction prototypes. These prototypes serve as references for target-domain recognition, where known-class nodes tend to exhibit prediction distributions similar to their corresponding source prototypes, while unknown-class nodes tend to show larger KL divergences, facilitating their identification.

This mechanism is particularly effective when known classes retain transferable semantic information across domains and the prediction distributions of known and unknown nodes exhibit sufficient separation. Furthermore, the margin visualizations in Figure 7 and the ablation results in Figure 3 further provide empirical support for the effectiveness of the virtual-class guided boundary learning.

We have further clarified this mechanism in Section 2.3 of the revised manuscript, highlighting the role of source-domain representations in target-domain unknown-class identification. We hope this explanation adequately addresses your concerns.

Q3: Discussion About the Impact of the Fixed η\eta-quantile to the Unknown-class Nodes.

R3: Thank you for your insightful question! Although the quantile parameter η\eta is fixed, the threshold δ\delta is dynamically determined by the fitted Beta distribution of KL divergence values in each target domain, allowing it to adapt to different divergence distributions. η\eta controls the strictness of unknown-class nodes identification. A smaller η\eta lowers the threshold and identifies more nodes as unknown, potentially improving unknown-class recall while increasing false positives. Conversely, a larger η\eta imposes a stricter threshold, reducing false positives while potentially missing some unknown nodes. Thus, when the proportion of unknown nodes varies across target domains, we set a moderate η\eta to control the trade-off between unknown-class precision and recall. Furthermore, in Section 3.4 we conduct sensitivity analyses with η\eta in 0.7,0.8,0.9,0.95,0.99\\{0.7,0.8,0.9,0.95,0.99\\}, as shown in Figure 4. The results indicate that η=0.9\eta=0.9 achieves a favorable balance in overall ACC and HS across the evaluated transfer tasks.

We hope this explanation adequately addresses your concerns.

Q4: Discussion About How to Adaptively Determine the Threshold.

R4: Thank you for your valuable question! We adopt a distribution-aware modeling strategy to derive an adaptive identification threshold rather than relying on a manually specified divergence threshold. The key insight is to decouple the selection criterion from the actual divergence threshold. The fixed quantile parameter η\eta specifies the proportion of nodes selected from the upper tail of the distribution, while the fitted Beta distribution determines the corresponding threshold δ\delta based on the underlying divergence distribution. Consequently, the threshold adapts to variations in target-domain data distributions while maintaining a consistent selection criterion.

Furthermore, we have clarified the rationale behind this distribution-aware thresholding strategy in Section 2.4 of the revised manuscript. We hope this explanation adequately addresses your concern.

Q5: Discussion about the Beta Distribution.

R5: Thank you for the comment! We conduct experiments to explore the effects of both GMMs and the Beta distribution for modeling the calculated KL divergence values. The experimental results (ACC/HS) are presented below:

TasksBeta Distribution ModelGaussian Mixture Model
A2C75.33 ±\pm 1.19 / 74.43 ±\pm 1.3570.22 ±\pm 0.99 / 64.17 ±\pm 0.99
A2D73.83 ±\pm 1.70 / 71.33 ±\pm 2.2669.08 ±\pm 1.89 / 60.75 ±\pm 2.96
C2D74.23 ±\pm 2.92 / 70.60 ±\pm 5.5269.49 ±\pm 2.46 / 59.63 ±\pm 4.66
C2A66.89 ±\pm 4.63 / 64.72 ±\pm 5.9060.80 ±\pm 3.97 / 53.17 ±\pm 4.70
D2A59.21 ±\pm 1.77 / 53.58 ±\pm 4.7154.33 ±\pm 1.73 / 44.58 ±\pm 4.59
D2C68.64 ±\pm 1.19 / 65.07 ±\pm 2.8161.98 ±\pm 0.92 / 53.91 ±\pm 1.60

Our experiments show that Beta distribution modeling consistently outperforms GMM-based modeling across all six transfer tasks, especially in HS. Through further analysis, we identify the underlying reasons for this discrepancy. First, the calculated KL divergence values are intrinsically bounded within a specific range, whereas GMMs have unbounded support over (−∞,+∞)(-\infty,+\infty). This mismatch in support may lead to suboptimal fitting. By normalizing the KL divergence values, we map them to [0,1][0,1], which precisely matches the natural bounded support of the Beta distribution. Second, in OSGDA, the presence of unknown categories inevitably introduces open-set noise. Under such noisy conditions, the EM algorithm used to optimize GMMs exhibits reduced stability and is prone to becoming trapped in poor local optima, preventing reliable fitting. In contrast, parameter estimation for the Beta distribution maintains high stability and robustness to such noise.

In summary, We adopt the Beta distribution for distribution modeling, aiming to reliably and consistently identify high-quality unknown-class nodes. We hope this explanation addresses the reviewer's concerns.

Q6: About the Edge Down-weighting Strategy.

R6: Thank you for the comment! Incorrectly identifying known-class nodes as unknown may reduce the contribution of useful neighborhood information during subsequent message passing. Therefore we incorporate nonzero edge down-weighting to provide tolerance for such misidentification. Specifically, in Eq. (14), we multiply the weights of edges connected to predicted unknown-class nodes by a small fixed value, while leaving the remaining edge weights unchanged. Thus, even when a known-class node is misidentified, its connections can still participate in message passing with reduced weights. This preserves information propagation while suppressing the influence of predicted unknown nodes, thereby partially mitigating the issue you raised. Furthermore, we also conduct experiments with different perturbation coefficient ΔAij\Delta A_{ij} for edges connected to predicted unknown nodes, with the results (ACC/HS) presented below.

ΔAij\Delta A_{ij}A2DA2CC2D
{0,1}71.26/70.3473.44/72.4671.58/71.02
{0.1,1}74.70/74.1075.68/74.7775.68/73.34
{0.2,1}74.69/74.1475.49/74.9575.73/73.22
{0.5,1}72.54/71.8374.53/73.2674.21/72.01
{0.8,1}70.68/70.0173.24/72.8971.30/70.88

The results show that small nonzero perturbation coefficients yield favorable performance, whereas setting the coefficient to 0 leads to a performance decline. This indicates that the issue the reviewer raised does arise. We therefore retain a small nonzero coefficient to reduce the impact of incorrectly identifying known-class nodes as unknown during training.

We thank the reviewer for the constructive feedback, which helpes us improve the quality of our work. All revisions mentioned above have been incorporated into the revised manuscript and marked in red, please refer to it for details. We hope that our responses satisfactorily address the reviewer’s concerns.

Reviewer_8yeKOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates the problem of open set graph domain adaptation and proposes a decision boundary aware framework named DEAR. It first employs virtual class learning to optimize the decision boundaries of known classes, then it identifies high confident known and unknown nodes through class relation mining. Finally, it introduces perturbations in cross-domain contrastive learning to implement knowledge transfer. Extensive experiments demonstrate its competitive performance.

Critical strengths
  1. The problem is well-motivated. This paper highlights the overlooked importance of source domain decision boundary robustness in existing methods, providing a clear motivation for the proposed framework.
  2. The proposed framework is well-structured and consists of three complementary modules, with a clear overall architecture and data flow.
  3. This paper leverages sourc -domain class prototypes and virtual class information to identify reliable known and unknown target nodes, facilitating subsequent self-training and knowledge transfer.
  4. The experiments are thorough. Performance on citation, ogbn-arxiv, and web datasets demonstrates that the proposed framework is effective across different graph scales and structures.
Critical weaknesses
  1. There are some typos. For example, figure 3 is ablation study, but the caption is inconsistent with the actual figure content.
  2. Eq.14 defines structural perturbation as Apt=At∧ΔAA_p^t=A^t \land \Delta A, while the subsequent description implements edge downweighting by assigning a weight of 0.2. The meaning of this operator and its implementation remain unclear.
  3. The source domain virtual class is constructed using node labels, but how virtual classes are generated for unlabeled target nodes and how final unknown-class predictions are obtained remain unclear.
  4. The method applies feature perturbation only to the source domain and structural perturbation only to the target domain, without explaining why both perturbations are not applied to both domains.
Questions for authors
  1. Could the authors clarify the meaning of the operator ∧\land in Eq.14 and how the weights of edges unrelated to predicted unknown nodes are handled?
  2. How is the virtual class of the target domain obtained and how are the unknown-class predictions obtained during the final inference stage?
  3. Why is KL divergence used to measure the discrepancy between target nodes and source-domain class prototypes? What advantages does it offer over prediction entropy or other similarity measures?
  4. The experiments only include the average results of known and unknown nodes. Could the authors give an evaluation of the quality of unknown-node identification?
2 comments

First of all, we'd like to thank you for your time and effort in reviewing our work. We will spare no efforts to address your concerns.

Q1: The Meaning of ∧\wedge and the Detailed Implementations of Structural Perturbation in Eq. (14).

R1: Thank you for the comment! In Eq. (14), ∧\wedge denotes element-wise multiplication. For the perturbation coefficient ΔAij\Delta A_{ij}, we set it to 0.2 for edges connected to predicted unknown-class nodes, while keeping it at 1 for edges that are not connected to any predicted unknown-class nodes.

This design is motivated by two considerations. First, due to the absence of ground-truth labels in the target domain, DEAR may incorrectly identify some known-class nodes as unknown. In such cases, directly reducing the associated edge weights to 0 would be overly aggressive and could completely eliminate useful information propagation. Instead, we adopt a small nonzero coefficient to moderately suppress the influence of predicted unknown-class nodes during message passing while retaining some tolerance for potential misidentification. Second, we conduct additional experiments with different ΔAij\Delta A_{ij} for edges connected to predicted unkonwn-nodes, and the results (ACC/HS) are presented below:

ΔAij\Delta A_{ij}A2DA2CC2D
{0,1}71.26/70.3473.44/72.4671.58/71.02
{0.1,1}74.70/74.1075.68/74.7775.68/73.34
{0.2,1}74.69/74.1475.49/74.9575.73/73.22
{0.5,1}72.54/71.8374.53/73.2674.21/72.01
{0.8,1}70.68/70.0173.24/72.8971.30/70.88

The results indicate that DEAR achieves relatively stable and favorable performance when the coefficient is small but nonzero. Therefore, we adopt a fixed perturbation coefficient of 0.2 to achieve simple and effective structural perturbation.

Given that ∧\wedge is commonly used to denote "AND" operation, which may cause ambiguity. To avoid this, we have replaced ∧\wedge with ⊙\odot in Eq. (14) in the revised manuscript. Additionally, we also clarify the meaning of this operator and provide a detailed explanation of the perturbation coefficient selection in Section 2.5, please refer to it for details.

We hope that these clarifications adequately address the reviewer's concerns.

Q2: Discussion about Wi,virtW^t_{i,vir} and the Predictions of Unknown-calss in the Final Inference Stage.

R2: Thank you for the comment! Sicne target domain nodes without ground-truth labels, we first obtain their pseudo-labels from the classifier's predictions, then we construct the corresponding virtual-class vectors following the same way as in Section 2.3.

Regarding unknown-class prediction, during training, we identify potential unknown-class nodes through class-relation mining and assign them an additional unknown-class pseudo-label for subsequent target-domain self-supervised training. This enables the classifier to recognize nodes outside the known-class distribution. During inference stage, the classifier directly produces predictions over ∣Cs∣+1|C^s|+1 classes, where the (∣Cs∣+1)(|C^s|+1)-th output represents the predicted probability of belonging to the unknown class, thereby enabling unknown-class identification.

To make these implementation details clearer, we have incorporated the above explanations into the revised manuscript. We hope these clarifications adequately address the reviewer's concerns.

Q3: Discussion about the Motivation on Using KL Divergence and Its Advantages.

R3: Thank you for your question! We adopt KL divergence because it measures discrepancies between class-probability distributions while incorporating source-domain class information. In contrast, prediction entropy only measures the uncertainty of individual predictions, without considering their consistency with source-domain class prototypes. Thus, even unknown-class nodes may exhibit low entropy when incorrectly assigned high-confidence known-class predictions. Instead, KL divergence can capture such distributional discrepancies by explicitly comparing target-node predictions with source-domain class prototypes.

Furthermore, compared with cosine similarity or Euclidean distance, KL divergence is specifically designed to measure differences between probability distributions, accounting for the relative probability mass assigned to different classes. Therefore, we adopt KL divergence to better exploit source-domain class information for unknown-class nodes identification.

We hope this explanation adequately addresses the reviewer's concerns.

Q4: Evaluation of the Quality of Unknown-class Node Identification.

R4: Thank you for the comment! In addition to the overall ACC and HS metrics, we also evaluate the unknown-class identification performance using unknown-class accuracy (ACCuACC_u), and the results are partially shown below:

ACCuACC_uA2DD2AA2CC2AC2DD2C
Average73.5070.7976.4769.6473.4580.20

More details are in Appendix E.1 (Table 2), please refer to it for details. These results directly evaluate DEAR's ability to identify unknown-class nodes under different unknown-class configurations, demonstrating its effectiveness in unknown-class recognition.

To make these results more intuitive, we have further supplemented results in Table 2 in Appendix E.1 of the revised manuscript. We hope these clarifications adequately address the reviewer's concerns.

Q5: About the Typos.

R5: Thank you for the suggestion! We have carefully proofread the entire manuscript, identified and corrected typographical errors and inaccurate figure captions, and incorporated these corrections into the revised manuscript. We sincerely appreciate your valuable feedback, which helps improve the overall quality of our paper.

Q6: Discussion about The Motivation for Applying Different Perturbations to the Source and Target Domains.

R6: Thank you for your comment! Since the source domain contains only known-class nodes, structural perturbation is unnecessary and may suppress useful topological information. Meanwhile, directly applying feature perturbation to the unlabeled target domain may increase uncertainty in pseudo-label learning. Therefore, applying both perturbation strategies to both domains may not necessarily yield additional benefits. So we adopt domain-specific perturbation strategies to address the primary challenges of each domain.

We have further clarified this design motivation in Section 2.5 of the revised manuscript. We hope this explanation adequately addresses the reviewer's concerns.

Thank you aganin for your valuable comments. All of the modifications mentioned above are highlighted in red in the revised manuscript, please refer to it for details. And we hope our responses can address your cocnerns.

Comments on the paper

7 comments

Thank you for your valuable comments regarding our work. We will make every effort to address your concerns.

Q1: Discussion about the Theoretical Interpretation.

R1: Thank you for your insightful comments about the theoretical analysis of the virtual class used in our DEAR regarding the margin and variance. We have carefully revisited both theoretical analyses and refined their corresponding claims to more accurately characterize the roles of virtual classes in our proposed method.

For Theorem 1, we conduct a more rigorous theoretical derivation, revising the original claim that "virtual classes can enlarge classification margins" to a more precise statement that "virtual classes can facilitate classification margin optimization." Specifically, by decomposing the Virtual Softmax loss and its gradients, we demonstrate that its gradient can be expressed as a weighted combination of two components: normal classification supervision and virtual class supervision, which are shown as follows: ∇Lvir,i=(1−αi)∇Lce,i⏟Normal supervision+αi∇divir⏟Virtual supervision,αi=elvir∑j=1∣Cs∣eWj⊤Fθ(Xi)+elvir.\nabla\mathcal{L}_{vir,i}=\underbrace{(1-\alpha_i)\nabla\mathcal{L}_{ce,i}}_{\text{Normal supervision}}+\underbrace{\alpha_i\nabla d_{i}^{vir}}_{\text{Virtual supervision}}, \quad \alpha_i=\frac{e^{l_{vir}}}{\sum_{j=1}^{|C_s|}e^{W_j^{\top}F_\theta(X_i)}+e^{l_{vir}}}.

The latter introduces an additional feature alignment signal that encourages node embeddings to align with the directions of their corresponding class centers, thereby providing complementary guidance for classification margin optimization. Furthermore, we derive a theoretical lower bound on the classification margin shown as:

m_i\geq\|F_\theta(X_i)\|g_{y_i} -D_{y_i}\sqrt{\frac{2\|F_\theta(X_i)\|d_i^{vir}}{\|W_{y_i}\|}}, \quad g_{y_i}=\min_{j\ne y_i}(W_{y_i}-W_j)^\top&\frac{W_{y_i}}{\|W_{y_i}\|}, \quad D_{y_i}=\max_{j\ne y_i}\|W_{y_i}-W_j\|, \end{aligned}$$ it is governed by the differences among class centers and the angular deviations between node embeddings and their corresponding class centers. Under certain conditions, this mechanism can further facilitate the enlargement of classification margins. For Theorem 2, which investigates the properties of virtual classes, we **retain the original variance property and introduce a complementary property of virtual class vectors,** which can be shown as: $$\underbrace{\mathcal{V}(l_{vir})\leq L_w^2 \cdot \mathcal{V}(\|F_\theta(X_i)\|)}_{\text{Direction Invariance}}, \quad \underbrace{\|W_{i,vir}-W_{j,vir}\|_2^2= 2\|W_c\|^2(1-\cos\phi_{ij})}_{\text{Scale Invariance}}.$$ Specifically, the variance property shows that the intra-class variance of virtual logits is determined by variations in node embedding norms and is not directly affected by directional changes, **exhibiting direction invariant intra-class consistency.** The newly introduced vector property further demonstrates that virtual-class vectors belonging to the same class share an identical norm, while their pairwise distances depend on the directional differences between node embeddings and remain invariant to positive rescaling of the embeddings, **exhibiting scale invariant intra-class consistency.** These two complementary properties characterize virtual classes from both scalar and vector perspectives, **providing statistical and geometric foundations for subsequent intra-class relation mining.** **We have updated both theorems, their corresponding proofs in the Appendix, and the relevant discussions throughout the manuscript to ensure consistency between the theoretical claims and their mathematical derivations.** Please refer to the revised manuscript for further details. We hope these revisions adequately address your concerns.

Q2: About the Selection Mechanism.

R2: Thank you for your valuable comments! In our response to Reviewer cHHZ Q1, we extend Theorem 2 by introducing a complementary property that characterizes the scale invariance of virtual-class vectors. Specifically, the distances between virtual-class vectors of nodes belonging to the same class depend only on their directional consistency, providing a theoretical basis for using virtual-class vectors in Eq. (12) to mine intra-class relationships and select high-confidence known-class nodes.

Regarding the generation of pseudo-labels and the identification of unknown-class nodes in the target domain, we first obtain target-node predictions from the classifier, then we compute the KL divergence between target-node predictions and source-class prototypes, model the resulting divergence distribution, and distinguish known-class and unknown-class nodes using a threshold. Subsequently, we assign unknown-class pseudo-labels to the identified unknown nodes, while employing virtual-class vectors to select high-confidence known-class nodes for subsequent self-supervised training.

To investigate the effectiveness of the selection mechanism in Eq. (12), we conduct experiments that evaluate the precision and coverage of target-domain nodes selected with and without the proposed selection mechanism within the same training epoch on six transfer tasks. The results (accuracy (%)/coverage rate(%)) are presented below:

A2CA2DC2DC2AD2AD2C
With Selection95.00/22.00100.00/7.0084.28/29.0090.08/12.995.68/39.594.27/62.5
Without Selection35.51/86.0032.44/85.0034.92/80.0037.02/78.4526.12/88.0031.43/89.00

As shown in the results, the proposed selection mechanism substantially improves the precision of selected nodes compared with the setting without selection, while maintaining a reasonable level of coverage. This demonstrates that our method can effectively identify a certain number of high-confidence known-class nodes.

We sincerely appreciate the reviewer's constructive comments, which helpes us improve the quality of our work. All the aforementioned revisions have been incorporated into the revised manuscript and highlighted in red. We hope that our responses adequately address the reviewer's concerns.

Reviewer_cHHZOpenReview ↗

Do DEAR’s Gains Hold Without Target Labels for Model Selection?

  1. In the released main.py, target_eval uses target ground-truth labels, and the training loop reports the epoch with the best target H-score. Were Tables 1–2 selected this way? What results hold when checkpoints are selected without target labels?

  2. Why does the released virtual_target compare scalar virtual logits, whereas Eq. (12) and the revised Theorem 2 concern virtual-class vectors? The code also reverses Eq. (6)'s KL direction. Which implementation produced the reported results, and can you release the matching version?

  3. Does Eq. (12) outperform confidence, entropy, or normalized-prototype filtering at matched coverage? How does comparing 100% precision/7% coverage with 32.44%/85% on A2D isolate the proposed selector rather than the effect of selecting fewer samples?

  4. How does Figure 7's unnormalized top-two logit gap establish a geometric boundary improvement, when logit rescaling increases that gap without changing decision regions? Do true-label, normalized or angular margins show the same benefit?

  5. What supports fixed ϵ=0.5\epsilon=0.5 and η=0.9\eta=0.9 across varying unknown-node proportions? Can you show sensitivity to ϵ\epsilon and controlled changes in unknown-node proportions, rather than only an η\eta sweep on fixed tasks?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.