- 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
-
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.
-
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.
-
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
-
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.
-
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.
-
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.
-
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
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.
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 , where is the share of transported source nodes that leave their angular cell (controlled by Eq. 10 and ) and is the class-conditional shift over the angular cells. The same-cell assumption is the case ; when it does not hold, the bound remains valid and loosens by , 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.
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.
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.