acceptodds
Under review as a conference paper at ICLR 2027

Open-Set Cross-Network Node Classification via OOD-Guided Selective Alignment

Abstract

Cross-network node classification transfers knowledge from a labeled source graph to an unlabeled target graph under distribution shifts. In open-set settings, aligning target nodes from unseen classes with known source classes can cause negative transfer. Existing open-set approaches primarily rely on prediction-related signals to guide adaptation, leaving distributional evidence insufficiently exploited for unknown identification. We propose OGSA, an out-of-distribution (OOD)-guided selective alignment framework that uses distributional deviation as complementary evidence for identifying unknown target nodes. Specifically, agreement between classifier predictions and prototype-initialized clustering provides initial target pseudo-labels. For nodes with conflicting predictions, structure-aware OOD scores provide additional evidence to identify potential unknown nodes. The refined assignments supervise target predictions and guide signed domain adversarial learning. Experiments on diverse graph benchmarks show that OGSA achieves competitive performance in both known class classification and unknown class recognition under open-set domain shifts.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Networks, Graph Domain Adaptation, Open-Set Node Classification, Out-of-Distribution Detection

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_bKbuOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates open-set cross-network node classification, where a model transfers knowledge from a labeled source graph to an unlabeled target graph containing both shared and unknown classes. The authors propose OGSA, an OOD-guided selective alignment framework that uses distributional evidence to identify unknown target nodes. The method integrates prediction-based screening, prototype-initialized clustering, and structure-aware OOD scoring to refine target pseudo-labels, which then guide target supervision and signed domain-adversarial alignment. The paper further derives an open-set target risk bound to characterize the trade-off between identifying unknown nodes and incorrectly rejecting known ones.

Experiments across 15 transfer tasks demonstrate consistent improvements over a broad set of graph OOD detection and domain adaptation baselines. The results are encouraging and support the effectiveness of selective alignment in open-set graph transfer. However, the robustness of the proposed approach under varying open-set conditions and the contribution of its individual components require further examination.

Critical strengths

S1. The paper addresses negative transfer caused by aligning unknown target nodes with known source classes. Incorporating distributional evidence beyond prediction confidence is a reasonable approach to improving unknown-node identification.

S2. OGSA achieves the best reported ACC and H-score across all 15 transfer tasks, outperforming competitive methods such as UAGA and GraphRTA. The consistent gains, particularly in unknown-class recognition, support the effectiveness of the proposed framework.

S3. The paper examines the effects of OOD refinement, signed domain alignment, alternative OOD scores, and auxiliary objectives. The analysis of pseudo-label quality during training also helps explain how the iterative refinement process affects performance.

Critical weaknesses

W1. The paper argues that unknown nodes can be confidently assigned to known classes despite substantial distributional differences. However, OOD refinement is restricted to candidates passing the initial prediction-based screening ((T=T_1\cap T_2), (\Omega=T\cap U)). Thus, confidently misclassified unknown nodes may be excluded regardless of their OOD scores. This appears inconsistent with the central motivation. The authors should quantify how often this occurs and examine whether OOD evidence can recover nodes outside the initial candidate set.

W2. OGSA builds on clustering and signed adversarial alignment closely related to UAGA, while incorporating OOD scoring from GEBM. Moreover, replacing the GEBM score with a simpler Mahalanobis distance yields a nearly identical average H-score. This raises questions about the necessity of the full OOD refinement design. A controlled comparison with UAGA augmented by the same OOD scores, using matched selection budgets and training settings, would better isolate OGSA's contribution.

Questions for authors
  1. Since (\Omega=T\cap U), unknown nodes excluded by the initial prediction-based screening cannot be recovered by OOD refinement. What proportion of confidently misclassified unknown nodes are excluded at this stage, and how does this affect unknown-class detection?

  2. How does OGSA compare with UAGA augmented by the same OOD scores under matched training settings and candidate budgets? This would help isolate the contribution of the proposed selection mechanism.

  3. How are the Top-k budget and other hyperparameters selected without labeled target data? How sensitive is performance to the proportion of unknown target nodes?

  4. How are the Gaussian statistics for the unknown class estimated when the source graph contains only known classes? Are unresolved target nodes included in ordinary domain alignment, and if so, how is potential negative transfer avoided?

3 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/3)

Thank you for the constructive and detailed review. Our responses to each point are provided below.


1. Unknown nodes excluded by the prediction-based screening. (W1/Q1)

Response: Thank you for this insightful comment. We agree that confidently misclassified unknown nodes can be excluded by the prediction-based screening regardless of their OOD scores. This restriction is a deliberate trade-off rather than an oversight. OOD scores alone may also flag known nodes that deviate from the source distribution due to domain shift, and requiring agreement with prediction-based evidence keeps such incorrect rejections under control. Moreover, the candidate set TT is recomputed at every epoch, so a node excluded in one epoch can still be identified later as the predictions evolve during adaptation. To quantify this effect, we take the top-kk target nodes by OOD score over all target nodes, denoted by UallU_{\text{all}}, and examine the high-OOD nodes that are not pseudo-labeled as unknown by OGSA, i.e., the excluded set X=Uall∖BX = U_{\text{all}} \setminus B, where BB denotes the nodes pseudo-labeled as unknown after OOD refinement. We report:

(1) the fraction of truly unknown nodes that fall in XX,

(2) the fraction of truly unknown nodes within XX,

(3) the fraction of truly known nodes that fall in XX, i.e., high-OOD known nodes protected from incorrect rejection,

(4) the fraction of unknown nodes in XX at the first adaptation epoch that are pseudo-labeled as unknown at the last epoch.

Results (%) on the five-class citation benchmarks are averaged over 5 seeds.

Task(1) Unknown excluded(2) Unknown share in XX(3) Known protected(4) Recovered
C1→A113.742.76.289.5
D1→A16.039.44.382.1
A1→C19.961.82.276.3
C1→D17.122.74.886.2
D1→C18.821.311.383.6
A1→D115.120.016.271.6
Avg.10.134.67.581.6

The results lead to three observations. First, only 10.1% of unknown nodes are excluded despite their high OOD scores on average. Second, truly unknown nodes account for only 34.6% of the excluded set on average, so admitting these nodes would incorrectly reject more known nodes than it recovers unknown ones. In exchange, 7.5% of known nodes with high OOD scores are protected from incorrect rejection. Third, 81.6% of the unknown nodes excluded at the first adaptation epoch are pseudo-labeled as unknown at the last epoch, showing that the exclusion is largely temporary and that OOD evidence does help recover nodes outside the initial candidate set as adaptation proceeds.

To further examine this trade-off under domain shifts, we apply controlled target-side perturbations on the five-class citation benchmarks. For attribute shift, a fraction of feature dimensions of each target known-class node are resampled from their empirical marginal distributions. For structure shift, a fraction of intra-class edges among target known-class nodes are rewired to nodes of other classes. We then compare the fraction of target known nodes wrongly flagged by the OOD score alone (UU) with that after intersecting with the prediction-based candidates (Ω=T∩U\Omega = T \cap U). Results (%) are averaged over 5 seeds and 6 transfer directions.

SettingKnown nodes in UUKnown nodes in Ω\Omega
No shift13.33.6
Attribute shift 0.115.63.9
Attribute shift 0.217.04.5
Attribute shift 0.319.24.8
Structure shift 0.116.74.3
Structure shift 0.219.04.8
Structure shift 0.320.95.6

As attribute or structure shifts become stronger, the fraction of known nodes wrongly flagged by the OOD score alone increases from 13.3% to 20.9%, whereas it remains below 5.6% after the prediction-based screening is applied. This shows that the advantage of the prediction-based screening becomes more evident under domain shifts, where more known nodes deviate from the source distribution. Therefore, the prediction-based screening has a limited impact on unknown-class detection while effectively preventing incorrect rejections of known nodes.

We will add this analysis to the revised version.

AuthorsOpenReview ↗

Official Comment by Authors (2/3)

2. Relation to UAGA and necessity of the OOD refinement design. (W2/Q2)

Response: This is a valuable point. We agree that the clustering and signed adversarial alignment in OGSA are closely related to UAGA and that the OOD score is adopted from GEBM. Our core contribution lies in the OOD-guided refinement. It uses distribution discrepancy as complementary evidence for identifying unknown nodes and combines it with prediction-based evidence through the intersection Ω=T∩U\Omega = T \cap U. To isolate this contribution, we follow your suggestion and augment UAGA with the same GEBM score under the same intersection rule. UAGA+GEBM is tuned per direction over the same hyperparameter ranges, including the candidate budget kk. The results (ACC / HS, %) on the five-class citation benchmarks are reported below.

MethodC1→A1D1→A1A1→C1C1→D1D1→C1A1→D1Avg.
UAGA60.91‾\underline{60.91} / 58.33‾\underline{58.33}60.86‾\underline{60.86} / 62.13‾\underline{62.13}62.23 / 59.1452.13 / 63.72‾\underline{63.72}60.41 / 52.2662.23 / 58.2959.80 / 58.98
UAGA+GEBM58.74 / 56.9559.76 / 59.5865.77‾\underline{65.77} / 67.28‾\underline{67.28}58.38‾\underline{58.38} / 61.9765.07‾\underline{65.07} / 69.18‾\underline{69.18}70.92‾\underline{70.92} / 71.01‾\underline{71.01}63.11‾\underline{63.11} / 65.16‾\underline{65.16}
OGSA71.98 / 75.3768.76 / 70.5381.43 / 80.5374.54 / 76.8775.94 / 77.0473.54 / 74.3674.37 / 75.78

Adding the GEBM score improves UAGA on average, which confirms that distributional deviation is useful evidence beyond our framework. However, UAGA+GEBM still falls clearly behind OGSA under the same score and selection rule. This shows that the gain of OGSA does not come from simply attaching an OOD score to an existing method. It comes from how the OOD evidence is integrated into adaptation, including how pseudo-labels are constructed and how unresolved nodes are excluded from alignment.

Regarding the Mahalanobis distance, we view its nearly identical performance as evidence for our motivation rather than against it. Even the simplest distance to source known classes is sufficient to bring a large improvement. This indicates that existing methods largely overlook distributional information, and that the benefit of OGSA does not depend on a particularly strong OOD detector. This is also why we directly adopt an existing OOD score instead of designing a tailored one. Developing OOD scores specific to open-set graph domain adaptation is a promising direction for future work.

To further justify the necessity of the full OOD refinement design, in particular the role of the intersection, we compare it with several alternative strategies for fusing the prediction-based score and the OOD score:

(1) Weighted sum: since the two scores lie on different scales, we consider three normalization schemes (min-max, z-score, and rank normalization) and combine the normalized scores by a weighted sum, with the weight searched over 0 to 1; (2) Product: the two scores are also normalized and then multiplied; (3) Rank average: the two scores are converted to ranks and averaged; (4) Union: the intersection in OGSA is replaced by the union, i.e., Ω=T∪U\Omega = T \cup U.

The results (ACC / HS, %) are summarized below.

ModelFive-class citationNine-class citationogbn-arxiv
OGSA74.37 / 75.7865.47 / 65.9153.56 / 55.20
w/o OOD refinement68.07 / 69.2955.77 / 48.0046.78 / 40.74
OOD refinement only68.91 / 71.5859.00 / 58.1344.02 / 44.24
Weighted sum71.23‾\underline{71.23} / 73.05‾\underline{73.05}64.62‾\underline{64.62} / 65.04‾\underline{65.04}52.07 / 53.90
Product70.57 / 72.4164.46 / 64.8953.06‾\underline{53.06} / 54.96‾\underline{54.96}
Rank average70.74 / 72.7262.30 / 62.3148.55 / 54.04
Union70.64 / 70.6361.00 / 61.0052.37 / 54.24

As shown in the table, all fusion strategies outperform both w/o OOD refinement and OOD refinement only, confirming that distributional deviation provides complementary evidence for unknown identification. Among them, the intersection achieves the best ACC and HS on all benchmark groups without additional tuning, while the alternative strategies require extra hyperparameters and show less stable performance across benchmarks. This is consistent with our theoretical analysis: incorporating OOD evidence helps recover missed unknown nodes, but known nodes that deviate from the source only due to domain shift may be incorrectly rejected. Strategies such as Union, which accept a node as unknown if either score flags it, are more prone to such errors, whereas the intersection limits them by requiring agreement between prediction-based and distributional evidence. We will add these comparisons and discussions in the revised version.

AuthorsOpenReview ↗

Official Comment by Authors (3/3)

3. Hyperparameter selection and sensitivity to the unknown proportion. (Q3)

Response: Thank you for the question. Following the common practice in open-set graph domain adaptation, the results of all methods are reported under their best hyperparameters on the target domain, so the comparison is conducted under the same protocol. More importantly, even without access to target labels for selection, OGSA remains stable over wide ranges of most hyperparameters, as shown in Appendix D.6. This indicates that the performance of OGSA does not rely on fine-grained tuning with target labels, and a reasonable default configuration is sufficient in practice.

To examine the sensitivity to the proportion of unknown target nodes, we keep only 75%, 50%, or 25% of the unknown target nodes on the five-class citation benchmarks, while using the same configurations as in the original setting without re-tuning. We compare OGSA with w/o OOD refinement and OOD refinement only. The results (ACC / HS, %) are reported below.

Unknown keptOGSAw/o OOD refinementOOD refinement only
100%74.37 / 75.7868.07 / 69.2968.91‾\underline{68.91} / 71.58‾\underline{71.58}
75%69.82 / 71.6957.08 / 47.1365.63‾\underline{65.63} / 65.05‾\underline{65.05}
50%64.78 / 66.4252.36 / 35.7862.10‾\underline{62.10} / 58.72‾\underline{58.72}
25%59.78 / 58.9647.18 / 32.8456.69‾\underline{56.69} / 49.91‾\underline{49.91}

As the proportion of unknown nodes decreases, the performance of all methods degrades, but OGSA consistently achieves the best ACC and HS. Removing OOD refinement leads to the largest degradation, since prediction-based signals alone fail to identify the few remaining unknown nodes. Using OOD refinement alone also degrades notably, since shifted known nodes dominate the high-OOD nodes when unknown nodes become rare and are thus incorrectly rejected. By requiring agreement between the two sources of evidence, OGSA avoids both issues and remains robust across different unknown proportions. We will add this result in the revised version.


4. Gaussian statistics for the unknown class and alignment of unresolved nodes. (Q4)

Response: Thank you for the careful reading. First, we do not estimate any Gaussian statistics for the unknown class. The GEBM score is fitted only on source nodes, where a class-conditional Gaussian is estimated for each known class. A target node that is far from all known-class Gaussians therefore receives a high energy and is regarded as potentially unknown. Second, unresolved target nodes are not included in the alignment of the encoder. They are neither aligned with nor pushed away from the source domain, which mitigates potential negative transfer caused by uncertain nodes. We will clarify both points in the revised version.


Thank you again for your constructive comments. We will incorporate these results into the revised version and are happy to address any further questions.

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

This paper studies open-set cross-network node classification and proposes an OOD-guided selective alignment framework named OGSA. The proposed method combines classifier predictions, prototype-initialized clustering, and OOD scores to identify unknown target nodes. The resulting unknown node assignments are then used to guide target supervision and domain alignment. Experimental results on several graph benchmarks demonstrate the effectiveness of the proposed method.

Critical strengths
  • The paper studies a meaningful problem in open-set GDA and explores the use of distributional information to improve unknown node identification.
  • Extensive experiments with different types of baselines demonstrate the effectiveness of the proposed framework. The paper also provides ablation studies to evaluate the contributions of different components.
Critical weaknesses

In Figure 1, the authors show that the distance gap between known and unknown target nodes becomes smaller or even reverses after domain adaptation. However, it is unclear why this observation indicates that existing methods fail to utilize distributional information. Moreover, a larger average distance gap does not necessarily mean that distributional distances can effectively distinguish known and unknown nodes.

Questions for authors
  • GraphRTA does not have a separate pre-training stage in its original framework. Why does Figure 1 divide its training process into pre-training and full adaptation stages?
  • In Section 4.4, the authors adopt a two-stage training procedure with pre-training followed by adaptation. Why is a separate pre-training stage necessary? Could unknown node identification and domain adaptation be performed jointly from the beginning of training? Although the ablation study includes a variant without pre-training, it would be helpful to further explain the advantage of the proposed two-stage design over joint training.
3 comments

Thank you for your careful reading and detailed questions. We clarify each point below.


1. Interpretation of Figure 1. (W1)

Response: We appreciate your close examination of Figure 1. We would like to clarify that it is intended as a motivating observation rather than evidence that existing methods are flawed. After pretraining, unknown target nodes deviate more from the source known-class distributions than known nodes within the same confidence groups. After adaptation, this gap becomes smaller or even reverses, which is expected for existing methods since domain adaptation typically pulls target representations toward the source domain. This observation motivates us to ask whether preserving this distributional gap and using it to assist unknown identification could be beneficial. OGSA is designed along this line: it uses distributional deviation to identify unknown nodes and pushes them away from the source domain through signed alignment. As shown in Figure 5, OGSA preserves the gap after adaptation in the lower-confidence groups, which are the main targets of OOD refinement.

We also agree that a larger average distance gap does not necessarily mean that distributional distances can effectively distinguish individual known and unknown nodes. This is exactly why OGSA does not use distributional evidence alone, but combines it with prediction-based evidence through the intersection-based selection. This design is supported by our experiments. Using OOD evidence alone (OOD refinement only in Table 3) is clearly inferior to OGSA, indicating that distributional distances alone are insufficient. In contrast, replacing the GEBM score with the Mahalanobis distance while keeping the combination with prediction-based evidence (Table 6) achieves performance comparable to OGSA, indicating that distributional distances provide useful complementary evidence. We will revise the description of Figure 1 in the introduction to clarify its motivating role.


2. Pretraining stage of GraphRTA in Figure 1. (Q1)

Response: Thank you for pointing this out. We would like to clarify that whether a method originally includes a separate pretraining stage does not affect our analysis, since the division in Figure 1 is introduced for analysis rather than to describe the original pipeline of each method. Specifically, we divide the training process to answer two questions. First, does distributional deviation provide evidence for distinguishing known and unknown target nodes before adaptation? Second, is this evidence preserved and exploited during adaptation? To answer the first question, we need a common reference point before adaptation, which is denoted by "Pretrain": all methods are trained for 30 epochs with labeled source nodes. To answer the second question, we compare it with "Final", which denotes the representations after the full adaptation of each method, with all adaptation stages of GraphRTA kept unchanged. As discussed in our response to W1, this comparison serves as a motivating observation: the distributional gap exists before adaptation but becomes smaller or even reverses after adaptation, which motivates OGSA to preserve and exploit this evidence for unknown identification. We will clarify this in the revised version.


3. Necessity of the two-stage training procedure. (Q2)

Response: This is a fair question. We adopt the pretraining stage because selective alignment requires identifying potentially unknown nodes before adaptation begins, but source supervision alone provides no direct training signal for the unknown class. Pretraining with source supervision and auxiliary unknown-class objectives provides reliable initial unknown predictions, which are then refined through clustering and OOD evidence during adaptation. Without this stage, the model has not yet learned to predict the unknown class when adaptation begins, so the nodes it initially flags as likely unknown are less reliable. Since these nodes initialize the unknown cluster and constrain OOD refinement, the resulting pseudo-labels become less accurate. This explains the performance drop of the variant without pretraining in Table 3. However, this does not mean that OGSA relies entirely on the accuracy of pretraining. As discussed in Section 5.3 (Robustness to Pseudo-Label Quality), although removing pretraining leads to highly inaccurate initial pseudo-labels, OGSA quickly recovers their quality during adaptation, correcting rather than amplifying pseudo-labeling errors. Therefore, pretraining provides a better starting point, while the refinement during adaptation keeps OGSA robust to imperfect initial predictions. We will further clarify the advantage of the two-stage design in the revised version.


Thank you again for your helpful questions. We will revise the corresponding descriptions and are happy to clarify anything that remains unclear.

Thank you for your responses, which address my main concerns. I have decided to raise my score to 3.

Thank you very much for your thoughtful feedback and raising your score. We will incorporate the clarifications and additional results from the discussion into the revised version.

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

This paper proposes OGSA, an OOD-guided selective alignment framework for open-set cross-network node classification. The key idea is to incorporate distributional evidence into unknown-node identification, complementing existing prediction-based approaches. Specifically, OGSA combines classifier predictions, prototype-initialized clustering, and structure-aware OOD scores to refine target pseudo-labels, which subsequently guide target supervision and signed domain adversarial alignment. Experiments on 15 cross-network transfer tasks demonstrate substantial improvements over existing methods.

Critical strengths
  1. Meaningful problem and motivation.
  2. Strong empirical performance.
  3. Relatively comprehensive experimental analysis.
Critical weaknesses
  1. The proposed framework primarily integrates existing techniques, including prototype-initialized clustering, graph OOD detection, and signed domain adversarial learning. The novelty of the overall framework and the necessity of its specific design choices remain unclear.
  2. Although the paper provides a target risk bound, the analysis mainly characterizes the intuitive trade-off between recovering unknown nodes and incorrectly rejecting known nodes. Its connection to the proposed methodological choices is insufficiently established.
  3. The experiments do not fully isolate the contribution of OOD evidence from other design choices. Moreover, the framework's ability to distinguish unknown classes from severe domain shifts within known classes requires further investigation.
Questions for authors
  1. The overall framework appears to be a relatively straightforward combination of existing techniques. Prototype-initialized clustering and signed domain adversarial learning are largely inherited from UAGA, while the structure-aware OOD scores are obtained from the existing GEBM framework. The main contribution therefore lies in incorporating OOD evidence into the pseudo-label refinement process. While this combination is empirically effective, the paper does not sufficiently justify why the proposed intersection-based selection strategy is necessary. For example, would simpler alternatives, such as a weighted combination of prediction uncertainty and OOD scores, achieve comparable performance? Additional comparisons with straightforward fusion strategies would help establish the methodological contribution.

  2. Limited theoretical insights and connection to the proposed method. Theorem 1 establishes an open-set target risk bound involving the recovery of unknown nodes and the incorrect rejection of known nodes. However, this trade-off is relatively intuitive, and the bound primarily provides a theoretical interpretation of the refinement process rather than guiding its design. In particular, the analysis does not explain why the proposed OOD scoring mechanism or intersection-based selection strategy should be preferred over alternative approaches. Moreover, the quantities associated with unknown recovery and known rejection depend on unavailable target labels, making their practical optimization unclear. The authors should clarify how the theoretical findings directly motivate the specific methodological components and training objectives.

  3. Distinguishing unknown classes from domain shifts. A fundamental challenge in applying OOD detection to graph domain adaptation is that distributional deviation may arise from either unknown classes or domain shifts within shared classes. A target node belonging to a known class may still exhibit a high OOD score under substantial attribute or structural shifts. Although OGSA combines prediction-based and OOD-based signals to mitigate this issue, it remains unclear how robust the approach is under different types and magnitudes of domain shifts. I would encourage the authors to investigate the effects of attribute shifts, structural shifts, and varying unknown-class proportions to better understand when OOD evidence is beneficial and when it may introduce incorrect rejection.

6 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/4)

We are grateful for your thoughtful and in-depth review. Our point-by-point responses are provided below.


1. The novelty of the overall framework and the necessity of its specific design choices remain unclear. (W1/Q1)

Response: Thank you for the careful reading. We agree that several modules in OGSA come from existing work, and that our main contribution lies in incorporating OOD evidence into pseudo-label refinement. Beyond this step itself, our key insight is that, while domain adaptation typically seeks to eliminate distribution discrepancy, in the open-set setting this discrepancy also serves as complementary evidence for identifying unknown target nodes. We further establish at which stage of adaptation and in what form this evidence should be introduced, where the key design choice is the intersection-based selection. To further justify its necessity, we follow your suggestion and compare it with several alternative strategies for fusing the prediction-based score and the OOD score:

  1. Weighted sum: since the two scores lie on different scales, we consider three normalization schemes (min-max, z-score, and rank normalization) and combine the normalized scores by a weighted sum, with the weight searched over 0 to 1;
  2. Product: the two scores are also normalized and then multiplied;
  3. Rank average: the two scores are converted to ranks and averaged;
  4. Union: the intersection in OGSA is replaced by the union, i.e., Ω=T∪U\Omega = T \cup U.

For each strategy and benchmark group, we report its best average HS. The results (ACC / HS, %) are summarized below.

ModelFive-class citationNine-class citationogbn-arxiv
OGSA74.37 / 75.7865.47 / 65.9153.56 / 55.20
w/o OOD refinement68.07 / 69.2955.77 / 48.0046.78 / 40.74
OOD refinement only68.91 / 71.5859.00 / 58.1344.02 / 44.24
Weighted sum71.23‾\underline{71.23} / 73.05‾\underline{73.05}64.62‾\underline{64.62} / 65.04‾\underline{65.04}52.07 / 53.90
Product70.57 / 72.4164.46 / 64.8953.06‾\underline{53.06} / 54.96‾\underline{54.96}
Rank average70.74 / 72.7262.30 / 62.3148.55 / 54.04
Union70.64 / 70.6361.00 / 61.0052.37 / 54.24

As shown in the table, all fusion strategies outperform both w/o OOD refinement and OOD refinement only, confirming that distributional deviation provides complementary evidence for unknown identification. Among them, OGSA achieves the best ACC and HS on all benchmark groups, while the alternative strategies require extra hyperparameters and show less stable performance across benchmarks. This is consistent with our Theorem 1: incorporating OOD evidence helps recover missed unknown nodes, but known nodes that deviate from the source only due to domain shift may be incorrectly rejected. Strategies such as Union are more prone to such errors, whereas the intersection limits them by requiring agreement between prediction-based and distributional evidence. We will add this comparison in the revised version.

AuthorsOpenReview ↗

Official Comment by Authors (2/4)

2. Connection between the theoretical analysis and the proposed method. (W2/Q2)

Response: Thank you for this insightful comment. We agree that Theorem 1 mainly characterizes the trade-off between recovering unknown nodes and incorrectly rejecting known nodes. To make the analysis directly guide the design, we derive a corollary that removes the dependence on target labels.

Setup. Recall from Definition 2 in Section 3.2 that Ωτ={i∈C:Ei≥τ}\Omega_\tau = \lbrace i \in \mathcal{C} : E_i \ge \tau\rbrace. We first define the fraction of target nodes selected as unknown as

$ \rho_t(\tau) = \Pr\nolimits_t(i \in \Omega_\tau) = \frac{|\Omega_\tau|}{|V_t|}. $

The incorrect rejections are governed by Pr⁡t(Ei≥τ∣Y~<K)\Pr_t(E_i \ge \tau \mid \widetilde{Y} < K), i.e., the false-positive rate of the OOD test on target known-class nodes. Since all source nodes belong to known classes, the same test applied to the source graph gives its false-positive rate on the source domain,

$ \Pr\nolimits_s(E_i \ge \tau) = \frac{\big|\lbrace i \in V_s : E_i \ge \tau\rbrace\big|}{|V_s|}, $

which can be computed without target labels. We measure the gap between the two by the discrepancy of OOD scores between source nodes and target known-class nodes,

$ \mathcal{D}_{E} = \sup_{\tau' \in \mathbb{R}} \big| \Pr\nolimits_s(E_i \ge \tau') - \Pr\nolimits_t(E_i \ge \tau' \mid \widetilde{Y} < K) \big|. $

Corollary 1. Under the conditions of Theorem 1,

$ R_t^{\text{os}}(F) \le (1-\pi_{\text{unk}})\big[R_s(F^{\text{kn}}) + \mathcal{D}_{\mathcal{H}} + \Lambda_{\text{kn}} + \alpha_0 + 2\mathcal{D}_{E}\big] + \pi_{\text{unk}}\beta_0 - \rho_t(\tau) + 2(1-\pi_{\text{unk}}) \Pr\nolimits_s(E_i \ge \tau) + \varepsilon_{\text{con}}. $

Proof. By the definitions of munk(τ)m_{\text{unk}}(\tau) and mkn(τ)m_{\text{kn}}(\tau) in Definition 2,

$ m_{\text{unk}}(\tau) = \Pr\nolimits_t(i \in \Omega_\tau, \widetilde{Y} = K) = \Pr\nolimits_t(i \in \Omega_\tau) - \Pr\nolimits_t(i \in \Omega_\tau, \widetilde{Y} < K) = \rho_t(\tau) - m_{\text{kn}}(\tau). $

Hence the refinement terms in Theorem 1 become mkn(τ)−munk(τ)=2mkn(τ)−ρt(τ).m_{\text{kn}}(\tau) - m_{\text{unk}}(\tau) = 2 m_{\text{kn}}(\tau) - \rho_t(\tau).

Since Ωτ⊆{i:Ei≥τ}\Omega_\tau \subseteq \lbrace i : E_i \ge \tau\rbrace, mkn(τ)=Pr⁡t(i∈Ωτ,Y~<K)≤Pr⁡t(Ei≥τ,Y~<K)=(1−πunk)Pr⁡t(Ei≥τ∣Y~<K).m_{\text{kn}}(\tau) = \Pr_t(i \in \Omega_\tau, \widetilde{Y} < K) \le \Pr_t(E_i \ge \tau, \widetilde{Y} < K) = (1-\pi_{\text{unk}}) \Pr_t(E_i \ge \tau \mid \widetilde{Y} < K).

By the definition of DE\mathcal{D}_{E},

$ \Pr\nolimits_t(E_i \ge \tau \mid \widetilde{Y} < K) \le \Pr\nolimits_s(E_i \ge \tau) + \big| \Pr\nolimits_t(E_i \ge \tau \mid \widetilde{Y} < K) - \Pr\nolimits_s(E_i \ge \tau) \big| \le \Pr\nolimits_s(E_i \ge \tau) + \mathcal{D}_{E}. $

Therefore, mkn(τ)≤(1−πunk)[Pr⁡s(Ei≥τ)+DE]m_{\text{kn}}(\tau) \le (1-\pi_{\text{unk}})\big[\Pr_s(E_i \ge \tau) + \mathcal{D}_{E}\big]. Substituting these relations into Theorem 1 yields Corollary 1.

Compared with Theorem 1, Corollary 1 shows that the intersection-based selection optimizes the bound by increasing the fraction of selected target nodes ρt(τ)\rho_t(\tau) while controlling incorrect rejections through the source false-positive rate Pr⁡s(Ei≥τ)\Pr_s(E_i \ge \tau), where the latter is particularly important when unknown nodes form a minority of the target graph. In contrast, union or score-level fusion may select nodes with Ei<τE_i < \tau, whose incorrect rejections cannot be controlled by Pr⁡s(Ei≥τ)\Pr_s(E_i \ge \tau). This is consistent with our additional experiments, where the intersection outperforms these fusion strategies. The remaining terms of the bound further guide the other components of OGSA. First, the bound requires reducing the shared-class discrepancies DH\mathcal{D}_{\mathcal{H}} and DE\mathcal{D}_{E}, which motivates signed domain adversarial learning: aligning only the inferred known nodes reduces both discrepancies, while pushing the inferred unknown nodes away from the source increases their OOD scores and preserves ρt(τ)\rho_t(\tau). Second, the bound requires reducing the disagreement εcon\varepsilon_{\text{con}} between model predictions and refined assignments, which motivates target supervision with the refined pseudo-labels in Lt\mathcal{L}_{\text{t}}. Finally, the initial error rates α0\alpha_0 and β0\beta_0 motivate pretraining with Lunk\mathcal{L}_{\text{unk}} and Ladv\mathcal{L}_{\text{adv}}, which provides reliable initial unknown predictions before refinement.

We will add this corollary with its proof to Section 3 and the Appendix in the revised version.

AuthorsOpenReview ↗

Official Comment by Authors (3/4)

3. Distinguishing unknown classes from domain shifts and isolating the contribution of OOD evidence. (W3/Q3)

Response: Thank you for raising this important point. We agree that a known-class node may also receive a high OOD score under severe domain shifts, which is exactly why OGSA does not use OOD evidence alone but combines it with prediction-based evidence through the intersection. To thoroughly examine this, we conduct controlled experiments on the five-class citation benchmarks with three types of target-side perturbations of increasing magnitudes:

(1) Attribute shift: for each target known-class node, a fraction (0.10.1, 0.20.2, 0.30.3) of feature dimensions are resampled from their empirical marginal distributions;

(2) Structure shift: a fraction (0.10.1, 0.20.2, 0.30.3) of intra-class edges among target known-class nodes are rewired to nodes of other classes;

(3) Unknown proportion: only 75%, 50%, or 25% of the unknown target nodes are kept.

We compare OGSA with w/o OOD refinement, OOD refinement only, and the best fusion strategy from our additional fusion experiments (Weighted sum). The results (ACC / HS, %) are summarized below.

SettingOGSAw/o OOD refinementOOD refinement onlyWeighted sum
No shift74.37 / 75.7868.07 / 69.2968.91 / 71.5871.23‾\underline{71.23} / 73.05‾\underline{73.05}
Attribute shift 0.172.03 / 73.2659.89 / 61.4864.68 / 64.4668.25‾\underline{68.25} / 69.73‾\underline{69.73}
Attribute shift 0.267.72 / 68.7655.59 / 56.8662.70 / 62.0564.89‾\underline{64.89} / 66.20‾\underline{66.20}
Attribute shift 0.363.79 / 61.9545.62 / 42.2658.87 / 56.0660.24‾\underline{60.24} / 58.98‾\underline{58.98}
Structure shift 0.166.41 / 64.8455.28 / 55.8460.28 / 59.9662.37‾\underline{62.37} / 63.67‾\underline{63.67}
Structure shift 0.260.14 / 57.7447.20 / 45.3651.34 / 47.2353.08‾\underline{53.08} / 50.56‾\underline{50.56}
Structure shift 0.353.38 / 48.4442.29 / 38.1546.89 / 42.23‾\underline{42.23}46.92‾\underline{46.92} / 41.48
Unknown kept 75%69.82 / 71.6957.08 / 47.1365.63 / 65.0567.20‾\underline{67.20} / 68.18‾\underline{68.18}
Unknown kept 50%64.78 / 66.4252.36 / 35.7862.10‾\underline{62.10} / 58.7260.82 / 62.00‾\underline{62.00}
Unknown kept 25%59.78 / 58.9647.18 / 32.8456.69 / 49.9158.88‾\underline{58.88} / 57.57‾\underline{57.57}

As the shifts become stronger, all methods degrade, but OGSA consistently achieves the best ACC and HS across all types and magnitudes of shifts. OOD refinement only degrades notably, especially under structure shifts, which directly reflects your concern: shifted known nodes also deviate from the source distribution and are mistaken for unknown ones. w/o OOD refinement degrades even more, since unknown nodes may receive confident known-class predictions under shifts and are then aligned with the source domain, whereas OOD evidence can still capture their deviation. Score-level fusion is generally the strongest alternative, but its gap to OGSA widens under structure shifts, since a shift-inflated OOD score can compensate for a low prediction-based score and cause known nodes to be selected. In contrast, OGSA selects a node only when both sources of evidence agree, so shifted known nodes supported only by the OOD score are excluded, while unknown nodes supported by both are still identified.

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

This work presents OGSA, an OOD-guided selective alignment framework, which adopts distribution deviation to detect unknown target nodes. It first yields target pseudo-labels via the consensus between classifier outputs and prototype-based clustering. For nodes with inconsistent predictions, structure-aware OOD scores help discover unknown candidates. The refined labels supervise target prediction and signed domain adversarial learning. Extensive graph benchmark experiments validate OGSA’s strong performance on both known-class classification and unknown-class recognition under open-set domain shifts.

Critical strengths
  1. The paper is well-written and well-organized, which makes it readable.
  2. The research topic is important for the graph-based learning community.
  3. The experiments are extensive to validate the effectiveness of the proposed model.
Critical weaknesses
  1. The authors don't compare their model with the SOTA proposed in 2026 in Table 2.
  2. Based on Figure 2, this figure doesn't present the key contributions, i.e., the novel modules, based on the ideas. The key ideas are not novel, in my view.
  3. Algorithms published in the past three years should be highlighted in section 2 related works.
  4. The source code of this paper is suggested to be open source if this paper is accepted.
Questions for authors
  1. The authors don't compare their model with the SOTA proposed in 2026 in Table 2.
  2. Based on Figure 2, this figure doesn't present the key contributions, i.e., the novel modules, based on the ideas. The key ideas are not novel, in my view.
  3. Algorithms published in the past three years should be highlighted in section 2 related works.
  4. The source code of this paper is suggested to be open source, if this paper is accepted.
1 comment
AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank you for your valuable comments. We address each concern below.


1. Comparison with recent methods proposed in 2026. (W1/Q1)

Response: Thank you for the suggestion. We have added four recent methods proposed in 2026, including two graph OOD detection methods, TIDE [1] and X-EviProbe [2], and two closed-set graph domain adaptation methods, ADAlign [3] and DiffGDA [4]. All methods are evaluated under the same open-set settings and protocol as in Table 2. We also identified two open-set graph domain adaptation methods published in 2026, DFGDN [5] and UCDA [6]. However, both of them lack publicly available implementations of their key components, which makes a faithful reproduction infeasible within the short rebuttal period, so they are not included in the comparison. The results (ACC / HS / AUC, %) are summarized below, where bold and underline indicate the best and second-best results, respectively.

CategoryMethodFive-class citationNine-class citationogbn-arxiv
Graph OOD DetectionTIDE58.62 / 60.64‾\underline{60.64} / 74.29‾\underline{74.29}55.90‾\underline{55.90} / 56.75‾\underline{56.75} / 68.87‾\underline{68.87}40.98 / 43.46 / 54.64
Graph OOD DetectionX-EviProbe55.16 / 57.75 / 73.4754.16 / 54.90 / 66.5736.61 / 38.93 / 53.37
Closed-set DAADAlign57.36 / 58.75 / 66.6152.30 / 53.01 / 61.5242.48‾\underline{42.48} / 47.01‾\underline{47.01} / 62.06‾\underline{62.06}
Closed-set DADiffGDA58.70‾\underline{58.70} / 58.94 / 65.4554.20 / 54.89 / 63.1441.86 / 46.49 / 61.83
OursOGSA74.37 / 75.78 / 93.1665.47 / 65.91 / 79.1753.56 / 55.20 / 77.19

OGSA consistently outperforms these recent methods on all benchmark groups and metrics. This is consistent with our analysis in Section 5.2: graph OOD detection methods identify unknown nodes without cross-domain alignment, so distributional deviations caused by domain shifts within shared classes are mistaken for unknown classes, while closed-set DA methods align all target nodes with the source domain, including unknown nodes, which leads to negative transfer and low AUC. In contrast, OGSA combines OOD evidence with prediction-based evidence to identify unknown nodes and selectively aligns only the transferable ones. We will add these results to Table 2 in the revised version.


2. Novelty of the key ideas and Figure 2. (W2/Q2)

Response: Thank you for the comment. We agree that individual components of OGSA are adapted from existing work, and we do not claim them as our contributions. Our core contribution lies in the OOD-guided refinement. It uses distribution discrepancy as complementary evidence for identifying unknown nodes and combines it with prediction-based evidence through the intersection Ω=T∩U\Omega = T \cap U. This design is supported by both our theoretical analysis and experiments. We agree that the current Figure 2 does not clearly present this contribution, and we will revise Figure 2 to explicitly illustrate the OOD-guided refinement in the revised version.


3. Highlighting recent algorithms in Section 2. (W3/Q3)

Response: This is a helpful suggestion for improving the related work. In the revised version, we will extend the current Section 2 with more recent algorithms, such as those proposed in 2026. Specifically, we will add TIDE [1] and X-EviProbe [2] for graph OOD detection, ADAlign [3] and DiffGDA [4] for graph domain adaptation, and DFGDN [5] and UCDA [6] for open-set graph domain adaptation. We will also discuss how OGSA differs from these recent methods.


4. Open-source code. (W4/Q4)

Response: We fully agree on the importance of reproducibility. We will release the complete source code upon acceptance to facilitate reproduction and future research.


References:

[1] Danny Wang, et al. (2026). What Information Matters? Graph Out-of-Distribution Detection via Tri-Component Information Decomposition. ICML.

[2] Chenghua Guo, et al. (2026). X-EviProbe: Post-hoc Parameter-Free Evidential Uncertainty Quantification for Frozen Graph Neural Networks. ICML.

[3] Wei Chen, et al. (2026). Learning Adaptive Distribution Alignment with Neural Characteristic Function for Graph Domain Adaptation. ICLR.

[4] Wei Chen, et al. (2026). Learning Structure-Semantic Evolution Trajectories for Graph Domain Adaptation. ICLR.

[5] S. Fu, et al. (2026). Domain-Adaptive Fuzzy Graph Diffusion Networks for Open-Set Cross-Domain Node Classification. IEEE Transactions on Fuzzy Systems, 34(6), 1744–1757.

[6] R. Song, et al. (2026). Unknown-aware Graph Consistent Domain Adaptation Network for Open-Set Node Classification. Pattern Recognition, 115208.


Thank you again for your valuable suggestions. We hope our responses have addressed your concerns and would be glad to provide further details if any questions remain.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.