acceptodds
Under review as a conference paper at ICLR 2027

Transferable Structural Concordance for Graph Foundation Models

Abstract

*Graph foundation models* (*GFMs*) aim to transfer knowledge across heterogeneous graph domains. Existing approaches largely organize cross-domain transfer at the graph level, treating each graph as a domain-specific structural unit and thereby overlooking substantial node-level structural heterogeneity. We characterize this heterogeneity through *structural concordance*, which measures the local agreement between a node's topological and feature-induced neighborhoods and reveals complementary high- and low-concordance regimes across graph domains. Building on this observation, we propose a *concordance-aware hierarchical coordination* framework for multi-domain GFM pre-training. Within each graph, two complementary coordinators specialize in high- and low-concordance regimes, while node-wise soft routing adapts their influence to each node's concordance level. Graph-specific meta coordinators further integrate concordance-conditioned information and mediate communication across source domains. This hierarchy enables node-adaptive cross-domain coordination while preserving the original graph topologies, following the principle of *coordinate without homogenizing*. We theoretically establish the informativeness, estimability, and representation stability of concordance-aware coordination. Experiments across diverse graph benchmarks consistently demonstrate improved unseen-domain transfer over existing GFMs.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Graph Foundation Models, Graph Pre-training, Cross-Domain Transfer

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 3 reviews, above 85% of the others.
Reviewer_ezH2OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a cross-domain graph pre-training, TSCGFM, which measure the agreement between each node's topological and semantic (feature-induced) neighborhoods by using structural concordance. Each node receives information from two coordinators (high and low), with the weights determined by its concordance score. And Meta-coordinators are used to share knowledge across different graphs. Experiments across diverse graph benchmarks consistently demonstrate improved unseen-domain over existing method.

Critical strengths

S1. The proposed method is clear and easy to follow.

S2. The experiments cover diverse graph benchmarks and useful ablation studies.

S3. The paper provides theoretical analysis to support the method.

Critical weaknesses

W1. Structural concordance is too abstract. Although figure 1 shows its distribution across datasets, more visual examples would help readers understand this concept. Its connection to the learned embeddings and cross-domain transfer is also unclear.

W2. The paper utilizes SVD to align the dimension of different datasets without any adaptation. This will probably make the feature spaces still unaligned, especially when transferring to an unseen domain.

W3. The paper gives little detail on the cost of constructing feature-induced neighborhoods. Section 3.2 connects coordinators across all pairs of domains. This suggests (O(M^2)) cross-domain edges, but Section 5 reports (O(M)).

Questions for authors

Refer to W1-W3.

5 comments

W1. Structural concordance is too abstract. Although figure 1 shows its distribution across datasets, more visual examples would help readers understand this concept. Its connection to the learned embeddings and cross-domain transfer is also unclear.

A1. We thank the reviewer for this valuable suggestion. To provide a more intuitive understanding of structural concordance and its role in cross-domain knowledge transfer, we provide an additional visual illustration (see Figure or Figure 2 in the revised manuscript) comparing three coordination mechanisms for multi-domain graph pre-training.

(1) Visual illustration of hierarchical coordination. The figure illustrates three different ways of organizing information exchange across graph domains. The left subfigure shows isolated pre-training, where graphs are processed independently without cross-domain connections or message passing. The middle subfigure illustrates GCOPE, which introduces graph-specific virtual coordinators to facilitate cross-domain communication. However, its coordination mechanism primarily operates at the graph level, without explicitly distinguishing nodes according to their local structural characteristics. In contrast, the right subfigure illustrates our proposed hierarchical coordination framework, which introduces complementary high- and low-concordance coordinators, node-wise adaptive routing, and global meta coordination. This design explicitly accounts for node-level structural heterogeneity within each graph, enabling more fine-grained cross-domain knowledge exchange rather than treating all nodes within a domain uniformly.

(2) How structural concordance facilitates cross-domain knowledge transfer. As illustrated in Figure 1, although the distributions of structural concordance differ across graph domains, each graph contains nodes exhibiting varying concordance levels, including relatively high- and low-concordance nodes. This observation motivates organizing cross-domain knowledge transfer according to local topology–feature agreement rather than graph identity alone.

Specifically, as shown in Figure or Figure 2 in the revised manuscript, for each graph mm, we introduce two virtual coordinators: a high-concordance coordinator cmhc_m^{h} and a low-concordance coordinator cmlc_m^{l}. Each node uu is connected to both coordinators, with routing weights determined by its structural concordance score sus_u. Nodes vv with higher concordance scores assign greater routing weights avha_v^h to cmhc_m^{h} and smaller weights avla_v^l to cmlc_m^{l}, whereas nodes with lower concordance scores exhibit the opposite routing preference. These concordance-conditioned connections allow nodes with different local topology–feature relationships to receive different mixtures of coordinator information, thereby influencing their learned representations.

Furthermore, the global meta coordinator connects the graph-specific coordinators across domains, facilitating information exchange between corresponding concordance regimes in different graphs. Through this hierarchical mechanism, nodes exhibiting similar concordance levels across heterogeneous domains can preferentially benefit from shared information, while graph-specific coordinators preserve domain-specific characteristics. Thus, structural concordance serves not merely as a descriptive graph statistic, but as an explicit routing signal that shapes representation learning and organizes cross-domain knowledge transfer.

W2. The paper utilizes SVD to align the dimension of different datasets without any adaptation. This will probably make the feature spaces still unaligned, especially when transferring to an unseen domain.

A2. We thank the reviewer for raising this important concern. We clarify that SVD is used as a preprocessing step to ensure dimensional compatibility across heterogeneous graphs, rather than to explicitly align their feature semantics or distributions. This is a common practice in multi-domain graph pre-training, where representative methods such as GCOPE, MDGFM, and SAMGPT also employ dimensionality reduction techniques, including SVD or PCA, to accommodate heterogeneous input feature dimensions.

Importantly, explicit feature distribution alignment is not necessarily required for transferable representation learning in multi-domain graph pre-training. Different graphs may exhibit substantially different feature semantics and structural characteristics, and directly enforcing similar feature or embedding distributions across domains does not necessarily establish semantic correspondence and may suppress useful domain-specific information. Instead of explicitly aligning feature distributions, TSCGFM facilitates cross-domain knowledge transfer through structural concordance and hierarchical coordination. Specifically, structural concordance is computed independently within each graph based on the agreement between topological and feature-induced neighborhoods, without requiring direct feature correspondence across domains. These node-level topology–feature relationships guide concordance-conditioned soft routing to complementary coordinators, while cross-domain coordinator connections and meta coordination facilitate information exchange across heterogeneous graphs. Together with the shared GNN encoder and joint self-supervised pre-training, this mechanism aims to learn transferable representations without requiring explicit alignment of feature or embedding distributions.

For an unseen target domain, following the transfer protocol of GCOPE, we apply the same dimensionality-unification procedure to the target graph and subsequently fine-tune the pre-trained GNN encoder using available target-domain labels. Although the target features may differ semantically from those of the source graphs, our framework aims to transfer knowledge of local topology–feature relationships learned across diverse source domains, rather than requiring explicit feature correspondence between source and target graphs.

W3. The paper gives little detail on the cost of constructing feature-induced neighborhoods. Section 3.2 connects coordinators across all pairs of domains. This suggests O(M2)O(M^2) cross-domain edges, but Section 5 reports O(M)O(M).

A3. We thank the reviewer for highlighting the need to clarify the computational cost of feature-induced neighborhood construction and the inconsistency in our analysis of cross-domain coordination complexity. We provide a detailed breakdown below.

(1) Feature-induced neighborhood construction. For each source graph GmG_m with Nm:=∣Vm∣N_m:=|V_m| nodes and aligned feature dimension dd, constructing the feature-induced neighborhoods involves two steps:

  • Pairwise similarity computation: Computing the cosine similarities between all pairs of nodes requires O(Nm2d)O(N_m^2d) time.
  • Degree-matched neighbor selection: For each node uu, we select the dm(u)d_m(u) most similar nodes from the remaining Nm−1N_m-1 nodes. Using linear-time selection, this requires O(Nm)O(N_m) time per node, resulting in O(Nm2)O(N_m^2) time across all nodes.

Therefore, the total construction complexity is O(Nm2d+Nm2)=O(Nm2d)O(N_m^2d+N_m^2)=O(N_m^2d) per graph, and O(d∑m=1MNm2)O(d\sum_{m=1}^{M}N_m^2) across all source graphs.

Importantly, feature-induced neighborhood construction is performed only once as a preprocessing step, before learning the GNN and coordinator node representations. The constructed neighborhoods remain fixed throughout the entire training process. Thus, this computational cost is incurred only once before model training and does not contribute to the computational complexity of subsequent training iterations.

(2) Graph augmentation and training complexity. We agree with the reviewer that the pairwise cross-domain coordinator connections introduce O(M2)O(M^2) edges, rather than O(M)O(M) as previously stated in Section 5. We clarify the graph augmentation complexity as follows.

  • Additional coordinator nodes: Each of the MM source graphs introduces two graph-specific coordinators and one meta coordinator, resulting in 3M3M additional nodes. Thus, the augmented graph contains ∣V′∣=N+3M|\mathcal V'|=N+3M nodes, where N=∑m=1M∣Vm∣N=\sum_{m=1}^{M}|V_m|.
  • Node-to-coordinator connections: Each original node connects to its two graph-specific coordinators, introducing 2N=O(N)2N=O(N) additional edges.
  • Cross-domain coordinator connections: For each pair of distinct source graphs, the corresponding high-concordance coordinators and low-concordance coordinators are connected, introducing two edges per graph pair. In addition, the corresponding meta coordinators are connected across source graphs, introducing one additional edge per pair. Therefore, the total number of cross-domain coordinator edges is 3(M2)=3M(M−1)2=O(M2)3\binom{M}{2}=\frac{3M(M-1)}{2}=O(M^2).
  • Meta-coordinator connections: Each graph-specific coordinator connects to its corresponding meta coordinator, introducing 2M2M additional edges.

Consequently, the augmented graph contains

\qquad |E'| =|E|+2N+3M(M-1)/2 =|E|+2N+1.5 M^2 - 1.5M,$$ where $|E|=\sum_{m=1}^{M}|E_m|$. Therefore, the augmented graph has $O(N+M)$ nodes and $O(| E|+N+M^2)$ edges. Since GNN training operates on the fixed augmented graph, its message-passing complexity depends on the augmented graph size, independently of the one-time feature-induced neighborhood construction. Although the additional coordinator connections introduce $O(N+M^2)$ edges, the quadratic term is negligible in our experimental settings, where $M^2\ll N$. Therefore, the coordination structure introduces only $O(N)$ additional edges under this condition, resulting in $O(N+M)$ nodes and $O(|E|+N)$ edges in the augmented graph, while maintaining efficient message passing. Figure 6 further compares the empirical running time and accuracy of TSCGFM with GCOPE, MDGFM, and SAMGPT, demonstrating that TSCGFM maintains comparable computational efficiency while achieving consistently higher accuracy. We have revised Sections 5 to explicitly correct the cross-domain coordination complexity from $O(M)$ to $O(M^2)$.
Reviewer_wtUmOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes TSCGFM, a multi-domain graph foundation model that addresses node-level structural heterogeneity through structural concordance, which measures the agreement between topological and feature-induced neighborhoods. The framework combines concordance-conditioned dual coordinators, node-wise soft routing, and hierarchical meta coordination to facilitate cross-domain knowledge transfer. The authors provide theoretical analyses of the approximation advantage, estimation reliability, and representation stability of the proposed mechanism. Experiments across diverse graph benchmarks demonstrate improvements in unseen-domain transfer.

Critical strengths
  1. Well-motivated perspective on cross-domain graph transfer. The paper highlights an important limitation of existing multi-domain GFMs, which largely organize transferable knowledge at the graph level while paying limited attention to node-level structural heterogeneity. Structural concordance offers an intuitive and interpretable measure for characterizing local topology–feature relationships across heterogeneous graph domains.

  2. Coherent methodological design with theoretical support. The dual coordinators, node-wise soft routing, and hierarchical meta coordination form a coherent framework. I particularly appreciate Theorem 4.1, which characterizes when concordance-conditioned routing reduces approximation error relative to a single coordinator, and the stability analysis linking routing estimation errors to changes in node representations.

  3. Strong and comprehensive empirical evaluation. The experiments cover diverse graph domains and demonstrate consistent improvements over representative GFM baselines. The comparisons with uniform and random routing are particularly useful for assessing the contribution of concordance-conditioned routing.

Critical weaknesses
  1. Although Figure 1 illustrates substantial node-level heterogeneity in structural concordance, the paper does not investigate how the transfer performance varies across different concordance regimes in unseen target domains. Such an analysis would help clarify whether the proposed pre-training framework provides consistent benefits to nodes with different local structural characteristics.

  2. Although the ablation studies demonstrate the benefits of concordance-conditioned routing over uniform and random routing, the paper does not discuss whether other node-level structural measures, such as node degree or local feature smoothness, could provide similarly informative routing signals. This makes it difficult to isolate the specific advantages of structural concordance.

  3. The paper provides a computational efficiency analysis and runtime comparisons, but does not explicitly examine how computational costs scale with the number of source domains. Since the proposed framework introduces cross-domain coordinator connections, this aspect deserves further clarification.

Questions for authors
  1. Can the authors provide a concordance-stratified evaluation of unseen-domain transfer performance? Specifically, how do the performance gains over existing GFMs vary across target-node groups with different structural concordance levels?

  2. How does structural concordance compare with simpler node-level measures as a routing signal? Could the authors provide a brief discussion or comparison using alternatives such as node degree or local feature smoothness?

  3. How does the computational overhead scale with the number of source domains? In particular, could the authors clarify the number of coordinator connections introduced by the hierarchical coordination mechanism and discuss its scalability as the number of source graphs increases?

4 comments

AW1 & A1. We thank the reviewer for this valuable suggestion. To investigate whether TSCGFM benefits nodes with different local structural characteristics, we conduct a concordance-stratified evaluation on the unseen target domains. Specifically, we divide target nodes into five groups based on their structural concordance scores. Group 1 contains all nodes with zero concordance, while the remaining nodes are sorted by their positive concordance scores and divided into four equally sized groups (Groups 2–5). We then compare TSCGFM with GCOPE, a representative GFM baseline that does not explicitly account for node-level structural heterogeneity. The table below reports the accuracy improvements (in percentage points) of TSCGFM over GCOPE within each group.

DatasetGroup 1 (Lowest)Group 2Group 3Group 4Group 5 (Highest)
Cora+1.46+3.33+4.96+2.36+3.65
Computers+5.21+11.18+9.76+9.09+6.74

TSCGFM consistently outperforms GCOPE across all five concordance groups on both datasets, with improvements ranging from 1.46 to 4.96 percentage points on Cora and from 5.21 to 11.18 percentage points on Computers. Importantly, the gains are not confined to a particular concordance regime, indicating that the benefits of concordance-aware pre-training extend to nodes with diverse local structural characteristics. These findings provide additional empirical support for the effectiveness of our framework in handling node-level structural heterogeneity during cross-domain transfer.

AW2 & A2: Comparison with Alternative Node-Level Routing Signals. We thank the reviewer for this constructive suggestion. To further investigate the specific advantages of structural concordance, we conduct additional ablation studies by replacing concordance-conditioned routing with two alternative node-level signals: (1) local feature smoothness, defined as the average feature similarity between a node and its topological neighbors, susmooth=1dm(u)∑v∈Nm(u)sim⁡(X~u,X~v)s_u^{\mathrm{smooth}}=\frac{1}{d_m(u)}\sum_{v\in\mathcal N_m(u)}\operatorname{sim}(\widetilde X_u,\widetilde X_v), and (2) node degree, dm(u)d_m(u). The results are summarized below, where values in parentheses indicate accuracy differences relative to TSCGFM (in percentage points).

MethodCoraCiteseerPubMedComputersPhoto
Feature Smoothness39.66 ± 2.7653.76 ± 3.9244.48 ± 1.9237.95 ± 3.7461.28 ± 1.17
Node Degree40.42 ± 4.6557.85 ± 4.0142.11 ± 2.2841.15 ± 5.0663.14 ± 1.83
TSCGFM (Ours)44.40 ± 2.5359.64 ± 2.8946.75 ± 0.7550.21 ± 1.9264.78 ± 3.32

Structural concordance consistently outperforms both alternative routing signals across all four datasets, with improvements of 2.27–12.26 percentage points over feature smoothness and 1.79–9.06 percentage points over node degree. Unlike node degree, which primarily captures local connectivity, or feature smoothness, which measures feature similarity along existing edges, structural concordance explicitly characterizes the agreement between topological and feature-induced neighborhoods. This distinction provides a more informative basis for organizing node-level coordination across heterogeneous graph domains. The additional results support the effectiveness of structural concordance as a routing signal beyond simpler local structural or feature-based measures.

AW3 & A3: Scalability of Hierarchical Coordination. We thank the reviewer for highlighting this point. We clarify the computational overhead of TSCGFM with respect to the number of source domains MM. As discussed in Section 5 (Computational Efficiency), given NN original nodes across MM source graphs, TSCGFM introduces two concordance-conditioned coordinators and one meta coordinator per graph, resulting in only 3M3M additional nodes, i.e., ∣V′∣=N+3M|V'|=N+3M.

The additional connections consist of three components: (1) node-to-coordinator connections, introducing 2N2N weighted edges; (2) within-domain connections between dual coordinators and their meta coordinator, introducing 2M2M edges; and (3) cross-domain connections among corresponding high-concordance, low-concordance, and meta coordinators, introducing 3(M2)3\binom{M}{2} edges under the undirected all-to-all construction. Therefore, the augmented graph has ∣E′∣=∣E∣+O(N)+O(M2)|E'|=|E|+O(N)+O(M^2).

Importantly, the quadratic term scales only with the number of source domains MM, rather than the total number of original nodes NN, as cross-domain connections are established exclusively among coordinators. In our experimental setting, M2≪NM^2\ll N, so the additional edge complexity is dominated by the O(N)O(N) node-to-coordinator connections. Furthermore, Figure 5 shows that TSCGFM achieves higher accuracy than GCOPE, SAMGPT, and MDGFM while maintaining comparable empirical runtime.

These analyses clarify that TSCGFM introduces linear node-level coordination overhead and quadratic cross-domain coordinator connectivity, providing a favorable efficiency–performance trade-off when the number of source domains is moderate.

Thank you for your detailed responses to my questions. My concerns have been satisfactorily addressed, and I am therefore inclined to increase my overall rating to clear accept.

We sincerely thank you for your positive feedback and thoughtful reconsideration. We are pleased that our responses have satisfactorily addressed your concerns, and we greatly appreciate your constructive comments and careful evaluation of our work.

Reviewer_4qWUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TSCGFM, a concordance-aware hierarchical coordination framework for cross-domain graph foundation models. The method introduces structural concordance to measure the agreement between topological and feature-induced neighborhoods. Based on this measure, dual coordinators and node-wise soft routing capture local structural heterogeneity, while meta coordinators enable cross-domain information exchange. The paper provides theoretical analyses and demonstrates improved transfer performance on multiple graph benchmarks.

Critical strengths

S1. The paper identifies an important limitation of graph-level coordination and highlights the need to consider node-level structural heterogeneity in cross-domain graph pre-training.

S2. The combination of structural concordance, dual coordinators, and hierarchical coordination provides a reasonable approach to cross-domain knowledge transfer.

S3. The method achieves competitive performance across multiple benchmarks.

Critical weaknesses

W1. The concordance score may be affected by graph size, node degree, and feature quality. Its comparability across heterogeneous domains requires further investigation.

W2. Structural concordance and virtual coordinators are closely related to existing ideas. Comparisons with alternative structural measures and adaptive routing methods would better establish the contribution.

W3. The theoretical results do not directly establish improved cross-domain generalization.

W4. The leave-one-graph-out setting may retain graphs from the same domain category during pre-training. A stricter leave-one-domain-out evaluation and statistical significance tests would strengthen the conclusions.

Questions for authors

Q1. Why are several related GFMs discussed in the paper, such as MDGPT, MDP-GNN, BRIDGE, and OMOG, excluded from the experimental comparisons? Could the authors include these baselines or clarify why direct comparisons are not feasible?

5 comments

W1. The concordance score may be affected by graph size, node degree, and feature quality. Its comparability across heterogeneous domains requires further investigation.

AW1. We thank the reviewer for raising this important concern. We acknowledge that structural concordance may be influenced by graph size, node degree, and feature quality. However, our framework does not assume that concordance scores follow identical distributions across heterogeneous domains. Instead, its cross-domain comparability relies on the following two properties.

First, structural concordance provides a consistently defined and normalized characterization of local topology–feature agreement. Specifically, it measures the proportion of a node's topological neighbors that also belong to its degree-matched feature-induced neighborhood. By normalizing the neighborhood overlap by node degree, the score is bounded within [0,1][0,1] and has a consistent interpretation across graphs of different sizes and degree distributions. Nevertheless, this normalization does not completely eliminate the influence of graph size, node degree, or feature quality.

Second, node-level structural heterogeneity is a shared phenomenon across graph domains, even when their concordance distributions differ. As illustrated in Figure 1, different graphs exhibit distinct concordance distributions, while nodes with varying concordance levels coexist within individual graphs. This observation suggests that concordance can serve as a common structural characterization across domains without requiring identical score distributions.

Building on these properties, TSCGFM leverages concordance-conditioned coordination to facilitate cross-domain knowledge transfer. Specifically, each node is softly connected to both high- and low-concordance coordinators, with edge weights determined by its concordance score. Nodes with higher concordance scores assign greater weights to high-concordance coordinators and smaller weights to low-concordance coordinators, while nodes with lower scores exhibit the opposite pattern. This enables the coordinators to capture complementary structural characteristics within each graph. Furthermore, cross-domain coordinator connections and global meta coordination facilitate knowledge exchange across heterogeneous graphs based on these concordance-conditioned representations, rather than relying solely on domain identity or requiring identical concordance distributions.

Therefore, although graph size, node degree, and feature quality may affect the numerical concordance scores, TSCGFM does not require these factors or the resulting score distributions to be identical across domains. Instead, it leverages a consistently defined structural measure and the shared presence of heterogeneous concordance levels to support cross-domain knowledge transfer.

W2. Structural concordance and virtual coordinators are closely related to existing ideas. Comparisons with alternative structural measures and adaptive routing methods would better establish the contribution.

AW2. We agree that structural concordance, virtual coordinators, and soft routing are individually related to existing techniques. Our contribution lies not in introducing these components independently, but in integrating them into a node-adaptive coordination mechanism that organizes cross-domain knowledge transfer according to topology–feature agreement rather than solely graph identity.

(1) Distinction from existing approaches. Existing GFMs primarily organize cross-domain knowledge through graph-level or domain-specific mechanisms, such as virtual coordinators in GCOPE and domain-specific prompts or tokens in MDGPT and SAMGPT. In contrast, TSCGFM explicitly characterizes node-level structural heterogeneity through structural concordance, which measures the agreement between topological and feature-induced neighborhoods.

The key novelty in our TSCGFM lies in how structural concordance directly governs node-specific interactions with complementary coordinators to organize cross-domain knowledge transfer. Specifically, each node is softly connected to both high- and low-concordance coordinators, with its concordance score determining their relative contributions to the node's representation. This enables nodes within the same graph to receive different mixtures of information according to their topology–feature agreement, rather than treating all nodes from the same domain uniformly. Furthermore, cross-domain coordinator connections and meta coordination facilitate knowledge exchange across heterogeneous graphs through these concordance-conditioned representations.

Thus, our contribution is not simply the combination of structural concordance, virtual coordinators, and soft routing, but a unified coordination mechanism that uses node-level topology–feature agreement to determine how information is organized and exchanged across domains.

(2) Comparison with alternative structural measures. To investigate whether the benefits arise specifically from structural concordance rather than alternative structural statistics, we conduct additional ablation experiments by replacing concordance with two alternative structural measures: (1) local feature similarity, defined as the average feature similarity between a node and its topological neighbors, susim=1dm(u)∑v∈Nm(u)sim⁡(X~u,X~v)s_u^{\mathrm{sim}}=\frac{1}{d_m(u)}\sum_{v\in\mathcal N_m(u)}\operatorname{sim}(\widetilde X_u,\widetilde X_v), and (2) node degree, sudeg=dm(u)s_u^{\mathrm{deg}}=d_m(u). All other components of TSCGFM are kept unchanged.

MethodCoraCiteseerPubMedComputersPhoto
Local Feature Similarity39.66 ± 2.7653.76 ± 3.9244.48 ± 1.9237.95 ± 3.7461.28 ± 1.17
Node Degree40.42 ± 4.6557.85 ± 4.0142.11 ± 2.2841.15 ± 5.0663.14 ± 1.83
Concordance (Our TSCGFM)44.40 ± 2.5359.64 ± 2.8946.75 ± 0.7550.21 ± 1.9264.78 ± 3.32

Structural concordance consistently outperforms both alternatives across all five datasets. One possible explanation for this improvement is its relative characterization of topology–feature agreement across heterogeneous domains. Unlike node degree, which measures neighborhood size, or local feature similarity, which relies on absolute feature similarity values, structural concordance measures the overlap between topological and feature-induced neighborhoods based on the relative ranking of feature similarities. This makes it less directly dependent on absolute similarity scales, which may vary substantially across domains. Such a relative characterization may provide a more suitable routing signal for coordinating nodes across graphs with heterogeneous feature and structural distributions.

(3) Comparison with alternative routing strategies. Our contribution lies in conditioning node-wise soft routing on structural concordance and integrating it with complementary high- and low-concordance coordinators. To further evaluate this design, we compare against two alternative routing strategies: (1) hard routing, which assigns each node exclusively to either the high- or low-concordance coordinator based on its concordance score; and (2) attention routing, which learns node-wise routing weights from the node embeddings and coordinator representations through scaled dot-product attention:

e_u^r=\frac{(W_qh_u)^\top(W_k \boldsymbol{c}_m^r)}{\sqrt d},$$ where $r\in\\{h,l\\}$, $h_u$ and $\boldsymbol{c}_m^r$ denote the embeddings of node $u$ and coordinator $c_m^r$, respectively, and $W_q$ and $W_k$ are learnable projection matrices. | Method | Cora | Citeseer | PubMed | Computers | Photo | |---|---|---|---|---|---| | Hard routing | 39.82±2.57 | 58.25±2.41 | 45.00±3.99 | 45.00±3.57 | 64.36±1.05 | | Attention routing | 40.93±3.41 | 39.69±4.28 | 46.38±3.75 | 43.52±2.56 | 62.03±1.23 | | Soft routing (Our TSCGFM) | **44.40±2.53** | **59.64±2.89** | **46.75±0.75** | **50.21±1.92** | **64.78±3.32** | As shown in the table, our concordance-conditioned soft routing consistently outperforms both alternatives across all five datasets. Compared with hard routing, the improvements suggest that allowing nodes to receive continuously varying contributions from both coordinators is more effective than enforcing binary assignments. Furthermore, although attention routing introduces additional learnable parameters to adaptively determine routing weights from node and coordinator representations, it does not necessarily yield better performance. These results suggest that explicitly incorporating structural concordance provides an informative inductive bias for coordinating nodes with heterogeneous local topology–feature relationships. Together, these comparisons support the effectiveness of combining topology–feature agreement-based routing with hierarchical node-adaptive coordination to capture node-level structural heterogeneity and facilitate cross-domain knowledge transfer.

W3. The theoretical results do not directly establish improved cross-domain generalization.

AW3. To address this concern, we provide an additional theoretical analysis of target-domain risk under cross-domain distribution shift. Specifically, we establish the following theorem showing that the proposed dual-coordinator architecture with concordance-conditioned soft routing yields a tighter upper bound on target-domain risk than a single-coordinator architecture.

Theorem: Target-Domain Risk Bound under Cross-Domain Distribution Shift

Let PSP_S and PTP_T denote the source and unseen target distributions over a common measurable space of node contexts and labels. Following Section 4.1, let q(v)∈Rdq(v)\in\mathbb R^d denote the ideal coordinator message for node vv and α⋆(v)∈[0,1]\alpha^{\star}(v)\in[0,1] the ideal population routing coefficient.

Let the coordinator anchors cH⋆c_H^{\star} and cL⋆c_L^{\star} be minimizer of the source-domain approximation errors Esoft\mathcal{E}_{\mathrm{soft}} and the coordinator c⋆c^{\star} be the minimizer of the errors Esingle\mathcal{E}_{\mathrm{single}} under the single coordinator case. Consider the population-optimal routing messages msoft(v)=α⋆(v)cH⋆+(1−α⋆(v))cL⋆m_{\mathrm{soft}}(v)=\alpha^{\star}(v)c_H^{\star}+(1-\alpha^{\star}(v))c_L^{\star} and msingle(v)=c⋆m_{\mathrm{single}}(v)=c^{\star},

Let hr(v)=g(v,mr(v))h_r(v)=g(v,m_r(v)) and h⋆(v)=g(v,q(v))h^{\star}(v)=g(v,q(v)) denote the routing-induced and ideal-message predictors, respectively.

Assume that the loss is LL-Lipschitz with respect to the coordinator message and that the squared approximation errors satisfy ∣q(v)−mr(v)∣22≤B\\|q(v)-m_r(v)\\|_2^2\leq B for both routing mechanisms and all node contexts in the source and target supports.

Define Er:=EPS[∣q(v)−mr(v)∣22]\mathcal{E}_{r}:=\mathbb E_{P_S}[\\|q(v)-m_r(v)\\|_2^2] and dTV(PS,PT)=sup⁡A∣PS(A)−PT(A)∣d_{\mathrm{TV}}(P_S,P_T)=\sup_A|P_S(A)-P_T(A)|.

Then, for r∈soft,singler\in\\{\mathrm{soft},\mathrm{single}\\},

RT(hr)≤RT(h⋆)+LEr+B dTV(PS,PT)=:Ur.R_T(h_r)\leq R_T(h^{\star})+L\sqrt{\mathcal E_r+B\,d_{\mathrm{TV}}(P_S,P_T)} =: U_r.

Furthermore, suppose Var⁡PS(α⋆(v))>0\operatorname{Var}_{P_S}(\alpha^{\star}(v))>0 and define Δapp=∣Cov⁡PS(α⋆(v),q(v))∣22Var⁡PS(α⋆(v))≥0\Delta_{\mathrm{app}}=\frac{\\|\operatorname{Cov}_{P_S}(\alpha^{\star}(v),q(v))\\|_2^2}{\operatorname{Var}_{P_S}(\alpha^{\star}(v))} \geq 0.

By Theorem 4.1, $\mathcal{E}_{\mathrm{soft}} = \mathcal{E}_{\mathrm{single}} -\Delta_{\mathrm{app}}$. Then Usoft≤UsingleU_{\mathrm{soft}}\leq U_{\mathrm{single}}, indicating that soft routing with two complementary coordinators yields a target-domain risk upper bound no larger than that of a single-coordinator architecture. The inequality is strict whenever Δapp>0\Delta_{\mathrm{app}} > 0.

Proof:

Step 1: Bounding target risk by routing approximation.

By the LL-Lipschitz continuity of the prediction loss with respect to the coordinator message,

RT(hr)−RT(h⋆)≤LEPT[∣mr(v)−q(v)∣2].R_T(h_r)-R_T(h^{\star})\leq L\mathbb E_{P_T}[\\|m_r(v)-q(v)\\|_2].

Applying the Cauchy--Schwarz inequality yields

\tag{A}$$ **Step 2: Controlling cross-domain discrepancy.** Let $e_r(v)=\\|q(v)-m_r(v)\\|_2^2$. Since $0\leq e_r(v)\leq B$, the layer-cake representation and the definition of total variation give $$\left|\mathbb E_{P_T}[e_r(v)]-\mathbb E_{P_S}[e_r(v)]\right|=\left|\int_0^B[P_T(e_r>t)-P_S(e_r>t)]\,dt\right|\leq B\,d_{\mathrm{TV}}(P_S,P_T).$$ Consequently, $$\mathbb E_{P_T}[e_r(v)]\leq\mathcal E_r+B\,d_{\mathrm{TV}}(P_S,P_T).$$ Combining this inequality with Equation (A), we obtain $$R_T(h_r)\leq R_T(h^{\star})+L\sqrt{\mathcal E_r+B\,d_{\mathrm{TV}}(P_S,P_T)}. \tag{B}$$ **Step 3: Comparing target-risk upper bounds.** Theorem 4.1 shows the approximation advantage of node-level adaptive routing over single coordinator, i.e., $$\mathcal E_{\mathrm{soft}}=\mathcal E_{\mathrm{single}}-\underbrace{\frac{\\|\operatorname{Cov}_{P_S}(\alpha^{\star}(v),q(v))\\|_2^2}{\operatorname{Var}_{P_S}(\alpha^{\star}(v))}}_{\Delta_{\mathrm{app}}\geq0}. \tag{Adv}$$ Substituting Equation (Adv) into Equation (B), the respective upper bounds are $$U_{\mathrm{soft}}=R_T(h^{\star})+L\sqrt{\mathcal E_{\mathrm{single}}-\Delta_{\mathrm{app}}+B\,d_{\mathrm{TV}}(P_S,P_T)},$$ $$U_{\mathrm{single}}=R_T(h^{\star})+L\sqrt{\mathcal E_{\mathrm{single}}+B\,d_{\mathrm{TV}}(P_S,P_T)}.$$ Since $\Delta_{\mathrm{app}}\geq 0$ and the square-root function is increasing, $U_{\mathrm{soft}}\leq U_{\mathrm{single}}$, with strict inequality whenever $\Delta_{\mathrm{app}}>0$ and $L>0$. This completes the proof.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.