acceptodds
Under review as a conference paper at ICLR 2027

Radially Dominant Hyperbolic Transport for Graph Domain Adaptation

Abstract

Leading solutions in graph domain adaptation treat domain shift as distribution alignment, learning domain-invariant representations by reducing the discrepancy between two graphs. Such methods transfer well across visual scenes, where CNN features decouple appearance from semantic structure. In a graph, however, a node's adjacency, neighborhood scale and branch membership are deeply entangled, so learning domain-invariant features by directly aligning distributions is often impractical. To address this, through an in-depth analysis of how nodes behave during transfer, we present Compass, which recasts graph domain adaptation as the fine-grained movement of individual nodes, termed transport, driven by a position-adaptive hyperbolic velocity field. Specifically, the transport has two properties. The first governs how far each node moves: hubs, rescaled strongly by the shift, and peripheral nodes, barely touched, each receive a correction matched to their structural position. The second governs which way it moves: along the radial direction that carries the domain gap and away from the angular direction that carries class, so trajectories of different classes stay apart, transported labels remain valid without target labels, and a few integration steps suffice. Compass reveals that graph adaptation has so far neglected the direction of movement, and the constraint it introduces is flexible, orthogonal to current top solutions and plug-and-play, bringing consistent gains on mainstream benchmarks.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
distribution shift, graph neural networks, transfer learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ZKdnOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Compass, a hyperbolic transport framework for unsupervised graph domain adaptation. Its central idea is to correct domain shift primarily through radial movement while preserving angular information associated with class identity. The method combines a hyperbolic encoder, a flow-matching objective, an angular regularizer, and a Busemann classification head. Experiments on six benchmarks report consistent improvements over existing methods, particularly for tail classes, together with reduced transport costs.

Critical strengths
  1. The distinction between radial structural variation and angular class information provides an intuitive motivation. The radial-displacement and radius–depth diagnostics offer empirical support for this design.

  2. The reported results cover several graph domains, including citation, social, and airport networks. Improvements in Macro-F1 on MAG and ogbn-arxiv suggest potential benefits for class-imbalanced adaptation.

  3. Adding the angular constraint and classification head improves both DiffGDA and ADAlign. These experiments suggest that the proposed components may be useful beyond the complete Compass framework.

Critical weaknesses
  1. The observation that adjacency, neighborhood scale, and branch membership are entangled does not, by itself, establish that direct distribution alignment is impractical. Moreover, Compass still relies on distributional coupling through optimal transport and flow matching. The contribution therefore appears incremental, combining established hyperbolic representations, flow matching, and angular classification with an angular regularizer. The authors should clarify the conceptual advance over the hyperbolic flow-matching approach cited in Section 2 and provide more direct evidence for the claimed limitations of existing alignment methods.

  2. Small pictograms, colored symbols, and graphical annotations are repeatedly interspersed with prose and mathematical definitions, particularly in the introduction and methodology. They interrupt the reading flow and make an already dense presentation harder to follow. Most of these visual elements should be confined to figures, with consistent textual terminology used throughout the main discussion.

  3. Although ogbn-arxiv provides a larger benchmark with approximately 169,000 nodes, evidence on substantially larger graphs remains limited. Experiments on larger protein interaction networks would help assess both scalability and generalization beyond the current domains. Molecular datasets could also provide a useful extension, although the authors would need to explain how the current node-level formulation transfers to graph-level prediction. Runtime and peak-memory measurements across increasing graph sizes would strengthen the efficiency claims.

  4. Preserving angular predictions does not establish semantic correctness after adaptation. Equation (10) controls whether transported source nodes remain within their original angular decision regions. This does not establish that corresponding target classes occupy the same regions. Likewise, predominantly radial displacement does not necessarily imply class-conditional alignment. The paper should state the assumptions connecting angular stability to target-label validity more explicitly and evaluate cases where domain shift also changes angular class structure.

Questions for authors

see Weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer ZKdn (Part 1/2)

We sincerely thank Reviewer ZKdn for the valuable comments. We are glad that the reviewer recognizes the intuitive motivation behind our radial–angular distinction, the empirical support from our diagnostics, and the usefulness of our components beyond the complete framework. All revisions are marked in red in the updated PDF.

Weakness 1: Novelty and Conceptual Advance

Thank you for this careful reading. To clarify, Compass is driven by a critical, previously unaddressed finding: graph domain shift has a dominant direction. Its conceptual advance is direction-constrained GDA, which shifts the paradigm from where representations land to how they move (L098–102). Our novelty is threefold:

  • New Problem Insight. Existing GDA methods constrain where representations arrive and leave free the direction along which each node moves (L155–156, L216–221). We show that this direction matters: 81% of the source-to-target displacement is radial (0.74–0.94 on all six benchmarks), while the class lies in the angle (Eq. 3, Table 1), a factorization that, to our knowledge, had not been tested on adaptation data (L106–109). We are also the first to identify and quantify path entanglement in GDA (Eq. 4, L126–128).

  • Specialized Technical Design. Unlike standard Riemannian flow matching, which treats all directions of the velocity alike, Compass turns the polar structure of the space into the constraint itself (L188–190). It combines a position-adaptive velocity field, a more expressive transformation than a one-step mapping (L021–023); an exact radial–angular split of the velocity at O(d) cost (Eq. 5, L249–250); the angular penalty L_ANG, which turns class preservation into a geometric constraint that needs no target label or pseudo-label (Eq. 7, Eq. 10, L119–125); and a Busemann head that reads the class from direction (Eq. 8).

  • Empirical Necessity. In single-factor controls that share the encoder and hyperparameters (Table 6; App. Table 16), replacing the hyperbolic geometry with a flat or a spherical one, removing L_ANG, or applying L_ANG in a random basis each costs 2.3 to 7.2 points on ArnetMiner, Airport and MAG, and raises the path entanglement from 0.14 to 0.33–0.47. The gain thus stems from constraining the radial direction in hyperbolic geometry rather than from the objective alone, and Compass is a necessary geometric advance rather than an incremental technical swap. The theory agrees: flat space divides the margin between neighboring classes by a power of the class count, whereas negative curvature subtracts only its logarithm (Eq. 9, L298–304), and the share of transported labels that leave their class cell is bounded by the angular action that L_ANG controls (Eq. 10).

On the distributional coupling. As the reviewer notes, Compass keeps the coupling through optimal transport (Eq. 6). Our claim is that alignment alone is insufficient on graphs, and the revised PDF words it accordingly (Abstract, L017–018; Sec. 1, L077–078).

Direct evidence on existing alignment. Adding the angular head and L_ANG to ADAlign, a one-shot alignment whose objective is left unchanged, and to DiffGDA improves both on every benchmark, by 1.4 and 1.6 points on the MAG tail (Table 2, L419–424). Constraining a direction that alignment leaves free helps, so the free direction has a cost.

Given the discovery of the radial dominance of graph shift and its targeted solution via direction-constrained hyperbolic transport, we believe that our work deserves to be shared with the community.

AuthorsOpenReview ↗

Response to Reviewer ZKdn (Part 2/2)

Weakness 2: Inline Pictograms and Colored Symbols

Thank you for this suggestion, which we have fully adopted: in the revised PDF, every inline pictogram and colored symbol is removed from the text, the captions and the tables.

Weakness 3: Scalability and Extension to Other Graph Types

Thank you for the constructive suggestion.

Larger graphs. Compass is linear in the size of the graph: adaptation costs O(K|V|d + |E|d), and the coupling is computed within minibatches of 4,096 nodes, so no dense |V|×|V| object is formed (L354–356, L918). On ogbn-arxiv, with 169,343 nodes, 1.17M edges and 40 classes (Table 7), Compass outperforms all six baselines that we ran, by +1.4 Macro-F1 overall and +2.2 on the tail classes (Table 5, L462–464).

PPI and molecular graphs. PPI is multi-label, whereas the current head assigns each node to one angular cell (Eq. 8); molecular graphs require graph-level prediction, where the transported object becomes a pooled graph representation. Both extensions keep the radial–angular split and L_ANG. The revised Limitations describe them and state that they are outside what we test (App. A.3, L790–793); we believe that they are promising directions for future work.

Weakness 4: From Angular Stability to Target-Label Validity

Thank you for this precise reading. Eq. (10) concerns the source labels that the transport carries; it reaches the target through one assumption, which the revised PDF now states in the main text, and whose consequence we measure. The assumption is that each class occupies the same angular cell in both graphs up to a change of radius. It is now stated right after Eq. (10), together with the fact that no bound on the target risk is given when it fails, as under concept shift (Sec. 3.4, L317–319; App. A.2, L758–765). Its consequence is measured on target accuracy: training the head on transported rather than untransported source nodes adds 2.5 / 3.7 / 3.1 points on ArnetMiner / Airport / MAG (Table 11), and removing L_ANG costs 2.3 / 4.1 / 4.9 points (Table 6). When the shift also changes the angular class structure, the radial share flags it before training and without target labels (App. A.1, L746–751), and Compass falls back to its Euclidean counterpart: within 0.3 points of Micro-F1 in a controlled attribute-only shift (ρ_rad = 0.21; App. H, L1173–1177). The revised PDF moves these conditions into the main text (Sec. 4.2, L411–415).

Thank you again for the insightful comments. If there are any remaining questions, we are more than happy to address them.

Thank you for the response. After carefully reading the rebuttal and the other reviewers’ comments, my concerns about whether angular preservation maintains target-class semantics remain unresolved. I also do not see additional experiments on other graph types, such as molecular graphs. I therefore retain my current score.

AuthorsOpenReview ↗

Follow-up Response to Reviewer ZKdn

We sincerely thank the reviewer for the time and for carefully reading our rebuttal and the other reviewers' comments.

Angular preservation and target-class semantics. We agree that preserving the angular cells of the source nodes does not by itself ensure target-class semantics, and that this link requires an assumption on the target classes. As Reviewer srd9 raised the same point, we have made the link explicit in our reply to Reviewer srd9: the target risk of the angular head satisfies ϵT(h)≤ϵ~(h)+η+Δang+TV(pS,pT)\epsilon_T(h) \le \tilde{\epsilon}(h) + \eta + \Delta_{\mathrm{ang}} + \mathrm{TV}(p_S, p_T), where η\eta is the share of transported source nodes that leave their angular cell (controlled by Eq. 10 and LANGL_{\mathrm{ANG}}) and Δang\Delta_{\mathrm{ang}} is the class-conditional shift over the angular cells. The same-cell assumption is the case Δang=0\Delta_{\mathrm{ang}} = 0; when it does not hold, the bound remains valid and loosens by Δang\Delta_{\mathrm{ang}}, while class-conditional shifts in the radius do not enter it. We do not claim robustness to shifts that alter the angular class structure, and we state this regime as outside the guarantee (Sec. 3.4, L317–319). We will add the bound and its proof to Sec. 3.4 and App. A.2.

Other graph types. We acknowledge that the current evaluation covers node-level adaptation only. Molecular graphs require graph-level prediction, where the transported object becomes a pooled graph representation; we were not able to complete this extension within the discussion period, and we describe it in the Limitations (App. A.3, L790–793) as an important direction for future work.

Thank you again for your time and constructive feedback.

Reviewer_srd9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Compass studies how to transfer a model trained on a labeled source graph to an unlabeled target graph. Its central hypothesis is that, in hyperbolic representations, the radius primarily captures differences in structural scale, while the angle primarily carries class information. Moving nodes mainly in the radial direction while suppressing angular changes can therefore reduce the domain gap while preserving class semantics.

The paper formulates graph domain adaptation as continuous, per-node transport, combining hyperbolic flow matching, an angular penalty, and a classifier that predicts classes from direction. It further establishes theoretical connections between angular displacement and label invalidation, path entanglement, and numerical integration error. The angular constraint can also be incorporated into existing methods to improve their performance.

Critical strengths

First, the paper offers a valuable perspective centered on directional constraints. By linking radial changes to structural domain shift and angular information to class semantics, it provides a clear modeling hypothesis for method design. This matters because its potential contribution goes beyond merely adding model components or improving benchmark scores.

Second, the method has reasonably strong empirical support, and its components are transferable to existing methods. The paper reports performance improvements across multiple benchmarks, ablation results, and gains from incorporating its components into DiffGDA and ADAlign. These results support the practical value of the core components and are important factors in favor of acceptance.

Critical weaknesses

W1. The proposed method assumes that domain shifts are predominantly radial, while class-discriminative information is primarily angular. Although supported by empirical observations, this assumption may not hold when domain shifts significantly alter community structures or class-related connectivity patterns.

W2. Equation (10) bounds the probability that transported source nodes cross angular class boundaries. However, preserving source class assignments does not necessarily ensure accurate target-domain predictions, particularly under class-conditional distribution shifts.

W3. The Busemann angular head determines class predictions solely from angular coordinates, potentially overlooking discriminative information encoded in radial coordinates, such as node degree, hierarchical depth, or structural scale.

Questions for authors

Q1. How does Compass perform when domain shifts primarily affect angular rather than radial coordinates, such as changes in community structure or class-related connectivity?

Q2. How does the source-label preservation guarantee in Equation (10) relate to target-domain classification performance, particularly under class-conditional distribution shifts?

Q3. How does Compass handle tasks where radial information, such as node degree or hierarchical depth, is important for distinguishing classes?

3 comments
AuthorsOpenReview ↗

Response to Reviewer srd9

We sincerely thank Reviewer srd9 for the valuable comments. We are glad that the reviewer recognizes the valuable perspective centered on directional constraints, the clear modeling hypothesis it provides for method design, and the transferability of our components to existing methods. All revisions are marked in red in the updated PDF.

Weakness 1 & Question 1: Scope of the Radial Premise

Thank you for this insightful question. Compass is sensitive to how much of the domain gap lies in the radius, and this quantity is measured before training and without target labels.

① The radial premise is measured on every benchmark. Graphs from different sources differ mainly in structural scale, which a hyperbolic representation places in the radius (L080–088). On all six benchmarks, 81% of the source-to-target displacement is radial (0.74–0.94; Eq. 3, Table 1, L106–109).

② The same measurement flags the shifts that the reviewer describes. A shift that alters community structure or class-related connectivity moves nodes across branches, i.e., changes their direction (L240–242), which lowers the radial share ρ_rad by the definition of Eq. (3). Both ρ_rad and δ need only the encoder, a coupling and the two graphs (App. A.1, L746–751).

③ In that regime, Compass falls back to its Euclidean counterpart rather than below it, since L_ANG penalizes angular motion rather than forbids it. In a controlled transfer where only the node attributes shift (ρ_rad = 0.21), Compass matches a Euclidean transport within 0.3 points of Micro-F1, at 1.4× the wall-clock (App. H, L1173–1177). When the shift is partly angular, the transport still closes the angular part of the gap: projecting out the angular velocity costs 1.0 and 1.2 points on ArnetMiner and MAG, but only 0.4 on the purely structural Airport (Table 11, L1087–1091).

We have refined the related descriptions in the updated PDF.

Weakness 2 & Question 2: From Eq. (10) to Target Accuracy

Thank you for this precise question. Eq. (10) bounds the label noise of the training set that the classifier sees, and it reaches target accuracy through one stated assumption, whose consequence we measure.

Theoretical analysis. Eq. (10) bounds the share of transported source nodes that leave their angular cell along the path (L306–315); since the head is trained on these nodes (Eq. 12, L351–352), this is the label noise of its training set. It extends to the target under one assumption: each class occupies the same angular cell in both graphs up to a change of radius (App. A.2, L758–765). As suggested, the revised PDF states this assumption right after Eq. (10), and states that no bound on the target risk is given when it fails, as under concept shift (Sec. 3.4, L317–319).

Empirical evidence. With the encoder, the head and every hyperparameter fixed, training the head on transported rather than untransported source nodes raises target accuracy by 2.5 / 3.7 / 3.1 points on ArnetMiner / Airport / MAG (Table 11), and removing L_ANG, the penalty that controls the bound, lowers it by 2.3 / 4.1 / 4.9 points (Table 6).

Label-proportion shift. A balanced coupling may pair a source node with a target node of another class. In polar coordinates such a pair is an angular displacement, which L_ANG suppresses instead of following (App. A.2, L765–768). Consistently, the gains concentrate on the tail classes: +2.2 Macro-F1 on the tail of both MAG and ogbn-arxiv, against +0.6 and +0.2 on the head (Table 5, L462–464).

Weakness 3 & Question 3: Radial Information in the Head

Thank you for the question. The head reads direction, and the only quantity kept out of the decision is the absolute radius, which is the coordinate that the shift moves. Airport is the direct test: its labels are levels of airport activity (L083) and its nodes carry no informative attributes (Table 7), so the class can only come from structure, and it is where Compass has its largest margins (+1.3 to +1.5 Micro-F1 on all six transfers, Table 4). The reason is that the scale of a network is domain-variant (94% of the displacement on Airport is radial, Table 1), whereas the connectivity pattern that places a node in the network hierarchy carries its class (L084–088). Where absolute scale both determines the class and is shared across domains, the head extends with a per-class radial term, which the revised PDF describes (App. A.4, L817–820).

Thank you again for the insightful comments. If there are any remaining questions, we are more than happy to address them.

Thank you for the clarification. However, the assumption that each class occupies the same angular cell across domains appears relatively strong and may not hold under substantial class-conditional distribution shifts. Moreover, the theoretical connection between the source-label preservation bound in Eq. (10) and the target-domain classification risk remains unclear. Therefore, my main concern remains unresolved, and I am inclined to maintain my original rating.

AuthorsOpenReview ↗

Follow-up Response to Reviewer srd9

We sincerely thank the reviewer for the time and the thoughtful follow-up. We agree that the link from Eq. (10) to the target risk deserves an explicit statement, and we are glad to provide it below.

① Eq. (10) enters the target risk through an explicit bound. The Busemann head hh reads only the angular cell of a node (Eq. 8). Let ϵ~(h)\tilde{\epsilon}(h) be its risk on the transported source nodes with their carried labels (the training objective, Eq. 12), η\eta the share of transported source nodes that end outside their source angular cell (at most the share bounded by Eq. 10), πS(⋅∣y)\pi_S(\cdot \mid y) and πT(⋅∣y)\pi_T(\cdot \mid y) the distributions of class-yy nodes over the angular cells in the source and target graphs, and pS,pTp_S, p_T the class proportions. Then

ϵT(h)≤ϵ~(h)+η+Δang+TV(pS,pT),Δang=∑ypS(y)TV(πS(⋅∣y),πT(⋅∣y)).\epsilon_T(h) \le \tilde{\epsilon}(h) + \eta + \Delta_{\mathrm{ang}} + \mathrm{TV}(p_S, p_T), \qquad \Delta_{\mathrm{ang}} = \sum_y p_S(y) \mathrm{TV}\big(\pi_S(\cdot \mid y), \pi_T(\cdot \mid y)\big).

Proof sketch. The 0-1 loss of hh depends only on the pair (cell, label), so ϵT(h)−ϵ~(h)\epsilon_T(h) - \tilde{\epsilon}(h) is at most the total variation between the two joint distributions of this pair, which splits into TV(pS,pT)\mathrm{TV}(p_S, p_T) (transport carries the labels) and a class-conditional term over the cells. Each transported node is coupled with its source node, so the class-conditional distance between transported and source cells is at most the probability that the pair falls in different cells, which sums to η\eta; the triangle inequality adds Δang\Delta_{\mathrm{ang}}. Eq. (10) thus enters the target risk additively through η\eta, which LANGL_{\mathrm{ANG}} controls.

② The same-cell assumption becomes a measurable term, and it allows class-conditional shift in the radius. We fully understand the concern, and we agree that this assumption may not hold under substantial class-conditional shifts that alter the angular class structure. Such a reliance is shared by unsupervised domain adaptation in general: as Ben-David et al. [1] show, a guarantee on the target domain requires some assumption that relates the source and target labeling. In this light, we have aimed to keep our assumption as weak as possible and to make its effect explicit when it does not hold. The assumption stated after Eq. (10) is the special case Δang=0\Delta_{\mathrm{ang}} = 0; when it does not hold, the bound remains valid and loosens by Δang\Delta_{\mathrm{ang}}. In addition, two properties keep its scope limited:

  • The radius does not appear in the bound. A class-conditional shift in the radius, e.g., one class changing its hierarchical depth or degree scale differently from another, costs nothing; only a shift that moves a class across angular cells does.
  • It is weaker than generalized label shift [2], which underlies guarantees for conditional alignment and requires pS(z∣y)=pT(z∣y)p_S(z \mid y) = p_T(z \mid y) for the full representation zz; ours requires it only for the angular cell.

③ The terms match the measured effects and the stated limits. Training the head on transported rather than untransported source nodes raises target accuracy by 2.5 / 3.7 / 3.1 points (Table 11), and removing LANGL_{\mathrm{ANG}}, which enlarges η\eta, lowers it by 2.3 / 4.1 / 4.9 points (Table 6). Under concept shift, Δang\Delta_{\mathrm{ang}} is large, which is the regime we state as outside the guarantee (Sec. 3.4, L317–319); the radial share flags it before training and without target labels (App. A.1), and Compass then matches its Euclidean counterpart within 0.3 points (App. H).

We will add this bound and its proof to Sec. 3.4 and App. A.2. We sincerely thank you again for your time and for raising this point, which makes the theoretical part of the paper more complete. We hope this clarifies the connection, and we would be happy to address any further questions.

[1] Shai Ben-David, Tyler Lu, Teresa Luu, and Dávid Pál. Impossibility theorems for domain adaptation. In AISTATS, 2010.

[2] Remi Tachet des Combes, Han Zhao, Yu-Xiang Wang, and Geoffrey J. Gordon. Domain adaptation with conditional distribution matching and generalized label shift. In NeurIPS, 2020.

Reviewer_a196OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Compass, a method for unsupervised graph domain adaptation. It learns a source-to-target transport mapping within hyperbolic space and employs radial/angular decomposition alongside angular regularization to minimize interference with class-related directions during the adaptation process. Experimental results on various graph benchmarks demonstrate strong performance.

Critical strengths
  1. The method offers a meaningful geometric perspective. By distinguishing between radial and angular directions, Compass not only examines whether domain distributions are aligned but also accounts for the influence of the transport direction on the category structure.
  2. The experimental results and ablation studies are quite thorough. The proposed method achieves competitive results across multiple graph benchmarks. Table 2 further demonstrates that incorporating the angular head and penalty into DiffGDA and ADAlign leads to performance improvements, which supports the broader applicability of this design.
Critical weaknesses
  1. The scope of the method’s applicability requires further clarification. As shown in Table 1, all the datasets evaluated exhibit a pronounced radial-dominant domain shift. While this supports the method’s core assumption, the current experiments do not yet provide direct validation for scenarios where the domain shift involves an angular-dominant shift, or where the graph structure lacks a clear hierarchical characteristic.
Questions for authors

When the domain shift is no longer primarily concentrated in the radial direction, might the angular constraint limit necessary adaptation? Could the authors elaborate further on the conditions under which Compass is applicable?

3 comments
AuthorsOpenReview ↗

Response to Reviewer a196

We sincerely thank Reviewer a196 for the valuable comments. We are glad that the reviewer recognizes the meaningful geometric perspective we offer, the thorough experiments and ablations, and the broader applicability of our design. All revisions are marked in red in the updated PDF.

Weakness 1 & Question: Scope of Applicability

Thank you for this insightful question. The radial design follows a precise diagnosis: graphs from different sources differ mainly in structural scale, which a hyperbolic representation places in the radius, while the direction carries the class (L080–088, L102–104). Compass therefore applies when the shift is mainly radial, and this can be checked before training and without target labels (Table 1).

① The radial premise is measured on every benchmark. On all six benchmarks, 81% of the displacement is radial (0.74–0.94) and all graphs are tree-like (δ ≤ 2.5) (Eq. 3, Table 1, L106–109). The radial share ρ_rad needs only the encoder and a coupling, and δ only the two graphs (App. A.1, L746–751), so this measurement is the applicability check that the reviewer asks for. It also locates the gain: the margin over the strongest baseline is largest on Airport, where ρ_rad is highest (0.94; +1.3 to +1.5 on all six transfers, Table 4).

② When the shift is not radial, Compass falls back to its Euclidean counterpart rather than below it, since L_ANG penalizes angular motion rather than forbids it. In a controlled transfer where only the node attributes shift (ρ_rad = 0.21), Compass matches a Euclidean transport within 0.3 points of Micro-F1, at 1.4× the wall-clock (App. H, L1173–1177); in this regime, one should lower λ2 or use the Euclidean transport (c = 0).

③ The direction constraint is useful beyond the full framework. Adding the angular head and L_ANG to DiffGDA and ADAlign improves both on every benchmark, by 1.6 and 1.4 points on the MAG tail (Table 2, L419–424).

We have refined the related descriptions in the updated PDF to eliminate any confusion.

Thank you again for the insightful comments. If there are any remaining questions, we are more than happy to address them.

Thanks for the authors' response, and I choose to maintain my score.

AuthorsOpenReview ↗

Thank you for the follow-up

We sincerely thank the reviewer for the time and for maintaining a positive assessment of our work. Your question on the scope of applicability has helped us state more clearly when Compass applies and how to check it before training (Sec. 4.2, L411–415), and we will incorporate these clarifications into the final manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their valuable comments and suggestions!

In short, we identify that graph domain shift has a dominant direction, which existing GDA methods leave unconstrained, and we propose Compass, a position-adaptive hyperbolic transport that corrects the structural gap along the radius while protecting class semantics in the angle. Compass attains the highest mean on all sixteen standard transfers and improves two recent methods as a plug-and-play module (Tables 2–4).

We have updated the manuscript with the following revisions based on the reviewers' suggestions. All revisions in the updated version are highlighted in red:

  1. Add more discussion and details:
    • Add a new paragraph, "When the radial premise holds", on when Compass applies and how to check it before training (a196-W1/Q, srd9-W1/Q1, ZKdn-W4; Sec. 4.2, L411–415).
    • State the assumption that connects Eq. 10 with the target labels (srd9-W2/Q2, ZKdn-W4; Sec. 3.4, L317–319).
    • Refine the claim on direct distribution alignment (ZKdn-W1; L017–018, L077–078).
    • Add a discussion of label-proportion shift and a radial extension of the head (srd9-W2/Q2, W3/Q3; App. A.2, L765–768; App. A.4, L817–820).
    • Extend the Limitations with graph-level and multi-label tasks, graphs far from tree-like, and shifts that rewire class-related connectivity (a196-W1/Q, srd9-W1/Q1, ZKdn-W3; App. A.3, L790–799).
  2. Improve the presentation:
    • Remove all inline pictograms and colored symbols (ZKdn-W2).

Thanks again for all the effort and time, and we look forward to further discussions if there are any more questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.