acceptodds
Under review as a conference paper at ICLR 2027

Federated Few-Shot Graph Domain Generalization with Non-Aligned Label Spaces

Abstract

Federated domain generalization (FDG) on graphs aims to collaboratively learn from decentralized source domains and generalize to unseen target domains without centralizing private graph data. Existing methods typically assume a predefined prediction task with globally aligned label spaces, limiting their applicability to heterogeneous graph domains with different class cardinalities and semantics. To address this limitation, we formulate **Federated Few-Shot Graph Domain Generalization with Non-Aligned Label Spaces**, where clients learn transferable model parameters without cross-domain label correspondence and predict query nodes in an unseen graph from only a few labeled context nodes. We further propose **FedDARC**, which first characterizes graph domains through the optimization responses of a common reference encoder, learns geometry-preserving domain states, and uses them to residually calibrate node representations. It then performs label-space-agnostic relational inference by constructing task-relative labels and aggregating query-to-context relational evidence, avoiding globally shared class semantics or a fixed-output classifier. At inference, the learned shared model remain fixed, while the target domain state and task-relative labels are derived from the labeled context nodes for prediction. Extensive experiments on graph datasets demonstrate the effectiveness of FedDARC in generalizing to unseen graph domains with non-aligned label spaces. Our code is available at https://anonymous.4open.science/r/fedgraph-0B15.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
federated learning, domain generalization

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_Z8NZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

Unlike existing federated graph domain generalization methods which mainly assume globally aligned label space, this paper addresses a practical and novel problem, i.e., Federated Few-Shot Graph Domain Generalization with Non-Aligned Label Spaces, where different clients do not share common label space and target classes may be unseen during training. This paper proposes a FedDARC method to address the problem, which combines optimization-response-based domain characterization, domain-state-guided residual calibration, and task-relative relational inference.

Critical strengths
  1. The paper addresses a practical and novel problem, i.e., Federated Few-Shot Graph Domain Generalization with Non-Aligned Label Spaces, where different clients do not share common label space and target classes may be unseen during training.
  2. The paper provides theoretical analyze to support the calibration-induced relational improvement.
  3. Experimental results demonstrate the effectiveness of the proposed method compared with multiple baselines.
Critical weaknesses
  1. This paper aims to address Federated Few-Shot Graph Domain Generalization with Non-Aligned Label Spaces. However, in the proposed model design, it is unclear which parts are actually devised for graph data, i.e., how the graph topological structures are utilized?
  2. In the addressed Federated Few-Shot Graph Domain Generalization problem, the target graph has a few accessible labels for each target class, thus, it is possible to directly conduct few-shot node classification solely on the target graph. What is the benefit of conducting Federated Few-Shot Graph Domain Generalization in such case? How much improvement can be obtained by transferring the knowledge from the source graphs, as compared to few-shot node classification solely on the target graph?
  3. The proposed method seems to be quite complex, combining multiple components, but the motivation of each component is not clear enough.
  4. It seems that the selected baselines are not designed for the scenario with Non-Aligned Label Spaces. It is confused how to adapt them to the task of Federated Few-Shot Graph Domain Generalization with Non-Aligned Label Spaces for fair comparison.
Questions for authors
  1. In the proposed method, which components are actually devised for graph data, i.e., how the graph topological structures of the source graphs and the target graph are utilized?
  2. In the addressed Federated Few-Shot Graph Domain Generalization problem, the target graph has a few accessible labels for each target class, thus, it is possible to directly conduct few-shot node classification solely on the target graph. What is the benefit of conducting Federated Few-Shot Graph Domain Generalization in such case? In experiments, how much improvement can be obtained by Federated Graph Domain Generalization, as compared to few-shot node classification directly on the target graph?
  3. In experiments, how to adapt the baselines which are not devised for the shared label space to the new task of Federated Few-Shot Graph Domain Generalization with Non-Aligned Label Spaces?
5 comments

We appreciate the reviewer's thoughtful comments and the opportunity to clarify the key aspects of our work.


W1 & Q1. Graph-Specific Design

Thank you for highlighting this ambiguity. We address this concern from two perspectives: why graph-specific considerations matter (Graph-Specific Challenges) and how graph topology is incorporated into FedDARC (Graph Encoding and Domain Calibration and Relational Classification).

  • Graph-Specific Challenges. Unlike independent samples, node representations in graphs depend on both node features and neighborhood structures. Consequently, cross-graph shifts in topology and feature distributions can alter message passing and learned representations, making transfer to unseen graphs challenging. Combined with non-aligned label spaces, these challenges motivate our domain-aware representation calibration and task-relative relational classification.

  • Graph Encoding. Our proposed FedDARC employs a two-layer GraphSAGE GNN encoder, which explicitly incorporates graph topology through neighborhood aggregation. Thus, node representations capture both node attributes and structural information, rather than treating nodes independently. More details are provided in Appendix C.1\color{#c05a18}\text{Appendix C.1}.

  • Domain Calibration and Relational Classification. Domain states are derived from the GNN's optimization responses, which are influenced by graph topology, node features, and supervision. These states guide representation calibration across heterogeneous graph domains. Relational classification then matches topology-aware query and context embeddings using task-relative labels, enabling prediction without globally aligned class semantics. Thus, graph structure influences both domain characterization and relational prediction, rather than being used only as an encoder input.

We will clarify these graph-specific design in the revised manuscript.


W2 & Q2. Federated Source-Graph Transfer

We explain the benefits of source-graph transfer and quantify them through a target-only comparison.

  • Benefits of Source-Graph Training. Source-graph training learns transferable node representations and query-to-context matching rules, improving prediction when only a few labeled target nodes are available. Table 4 \color{#c05a18}\text{Table 4 } in the main paper further shows that performance improves as more source graphs are used for training.

  • Comparison with Target-Only Classification. The following Table 1 compares our method with target-only few-shot classification under the same target-data settings. It can be seen that our method improves accuracy by 19.20/18.33 points on Computers. These results support the value of transferring source-graph knowledge for few-shot prediction on unseen target graphs.

Table 1. Comparison with Target-Only Few-Shot Classification

DatasetCora|ogbn|Computers
1-shot3-shot|1-shot3-shot|1-shot3-shot
Ours45.4554.85|21.0823.72|39.5344.56
Target-only28.6638.03|8.3811.55|20.3326.23
Δ\Delta+16.79+16.82|+12.70+12.17|+19.20+18.33

W3. Clarification of the Components

FedDARC combines two complementary components to address cross-graph distribution shifts and non-aligned label spaces.

  • Domain-state-guided representation calibration. Cross-graph shifts can affect node representations and weaken relational matching. We therefore derive domain states from a common reference encoder's optimization responses and use them to guide residual calibration. This enables domain-dependent adjustments while retaining the original node representations.

  • Label-space-agnostic relational classification. Representation calibration alone cannot resolve differences in class cardinalities and semantics. We therefore use task-relative labels and query-to-context matching instead of a fixed-output classifier, enabling prediction without cross-domain label correspondence.

Supporting Evidence. Table 2 \color{#c05a18}\text{Table 2 } shows that removing domain calibration reduces Cora 5-shot accuracy from 58.20% to 54.13%, while removing label-relative inference further reduces accuracy to 13.04%, supporting the importance of task-relative prediction. Moreover, Proposition 1\color{#c05a18}\text{Proposition 1} characterizes the conditions under which calibration improves relational predictions, with supporting empirical analysis in Table 3 \color{#c05a18}\text{Table 3 }.

These results support the complementary roles of domain characterization, representation calibration, and label-space-agnostic relational classification.

W4 & Q3. Baseline Adaptation and Fair Comparison

Thank you for raising this concern.

  • Baseline Adaptation. Most existing FL/FDG baselines were originally designed for aligned label spaces, making their fixed-output classifiers unsuitable for direct transfer across domains with different class cardinalities and semantics. As described in Appendix C.1\color{#c05a18}\text{Appendix C.1}, our evaluation uses their adapted versions. Baselines requiring target-specific adaptation are further trained using only the few labeled target context nodes to accommodate the target label space, without accessing target query labels. In contrast, FedDARC performs context-conditioned relational prediction with fixed learned parameters, without target-specific fine-tuning.

  • Fair Comparison. All methods use the same source-domain training data, target context–query splits, and label budgets. We also maintain approximately equal total federated communication budgets, adjusting the number of communication rounds according to each method's per-round cost, as detailed in Appendix C.5\color{#c05a18}\text{Appendix C.5}. Target-specific adaptation does not involve additional client–server communication.

These protocols ensure consistent data access, target supervision, and federated communication constraints across methods, while accommodating their different prediction mechanisms under non-aligned label spaces.

Esteemed Reviewer,

We sincerely thank you for the time and effort you have devoted to evaluating our work.

We hope that our responses and revisions have adequately addressed your questions and concerns. If so, we would be grateful if you could kindly consider whether the current revisions might merit a reassessment of the score. Of course, should any technical quality points remain unclear or if you have further suggestions, we would be most happy to hear from you.

Best regards,

Authors

Reviewer_u4EKOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies Federated Few-Shot Graph Domain Generalization with Non-Aligned Label Spaces, where source clients hold graphs with different features, structures, and label spaces, and the shared model must classify query nodes in an unseen target graph from only a few labeled context nodes per class. The authors propose FedDARC, which learns a domain state from the optimization response of a common reference encoder to calibrate node representations, and predicts labels via query-to-context relational inference with task-relative labels. They provide a theoretical analysis of the calibration and of generalization to unseen domains, with experiments on four unseen target graphs.

Critical strengths
  1. This paper formulates a novel and well-motivated problem setting. FL with heterogeneous label spaces is a limitation of existing FDG work.
  2. The paper is clearly written and has a comprehensive empirical study. The paper comparisons against 11 baselines on four unseen domain. The authors connect the domain theory to measurable quantities.
Critical weaknesses
  1. The relative contribution of the domain state could be clarified. This lacks a clear description of how the baselines were adapted to non-aligned label spaces, add two key comparisons: FedAvg with the same head (w/o domain state) and a centralized upper bound.

  2. The paper does not report the inference time overhead. At test time, FedDARC run reference-encoder on the target context nodes to calculate the domain state, and this cost is not reported.

Questions for authors
  1. How did the fixed-output baselines handle clients with different numbers of classes?
  2. Baselines are mainly test on leave one source out setting. Can you provide leave one source out results compared with others?
3 comments

We appreciate the reviewer's positive assessment and insightful suggestions for strengthening our evaluation.


W1. Contribution of Domain Calibration and Additional Comparisons

  • Role of Domain Calibration. Although task-relative relational classification enables prediction without aligned label spaces, cross-graph distribution shifts can still affect node representations and weaken query-to-context matching. Domain-state-guided calibration addresses this issue by adjusting node representations using domain states derived from a common reference encoder's optimization responses.

  • Additional Comparisons. FedAvg with the same prediction head and without the domain state corresponds to w/o Domain Calibration in Table 2\color{#c05a18}\text{Table 2} of the main paper. This comparison retains the same graph encoder and relational prediction mechanism, allowing us to assess the contribution of domain calibration beyond the shared prediction head. We additionally include a centralized training reference to assess the performance gap between federated and centralized learning. As shown in the following table, domain calibration improves average accuracy from 36.6% to 39.9% across nine settings. These results support the contribution of domain calibration and show that FedDARC narrows the gap to centralized training compared with the uncalibrated variant.

MethodCora|ogbn-Products|Computers|Avg.
1-shot3-shot5-shot|1-shot3-shot5-shot|1-shot3-shot5-shot|
Centralized51.760.561.8|22.023.126.1|44.549.153.3|43.5
w/o Domain Calibration42.351.254.1|16.819.424.2|31.444.645.7|36.6
Ours45.554.958.2|21.123.724.9|39.544.647.0|39.9

W2. Inference Time

Thank you for pointing this out. We additionally report inference time under the 1-, 3-, and 5-shot settings. The reported time(ms) includes reference-encoder computation for domain-state estimation and subsequent query prediction.

Dataset1-shot3-shot5-shot
Cora4.614.574.77
Computers31.8632.0232.66
ogbn-Products115.87183.76256.63

The results show that inference time depends on the target graph and shot setting, with larger costs observed on ogbn-Products. These measurements explicitly account for domain-state estimation rather than reporting query prediction time alone.


Q1 Baseline Adaptation

  • Non-Aligned Label Spaces. Existing FL/FDG baselines generally assume aligned class semantics and fixed-output classifiers, making direct application to our setting inappropriate. We therefore evaluate their adapted versions rather than assuming a globally shared label space.

  • Adaptation and Fair Comparison. As described in Appendix C.1 and C.5\color{#c05a18}\text{Appendix C.1 and C.5}, baselines requiring target-specific adaptation are further trained using only the few labeled target context nodes to accommodate the target label space. All methods use the same source-domain training data, target context–query splits, and label budgets, without accessing target query labels.

Federated methods are also evaluated under approximately equal total communication budgets, with communication rounds adjusted according to per-round costs. These protocols provide consistent supervision and communication constraints while accommodating the different prediction mechanisms.


Q2 Leave-one-domain

We additionally conduct leave-one-domain-out evaluation on the four source graphs used in Table 1\color{#c05a18}\text{Table 1} holding out each graph as the target and using the remaining three for federated training. The following table shows that FedDARC outperforms FedAvg across all nine reported settings, with gains of 1.4–14.1 percentage points. These results provide further evidence that FedDARC can generalize to held-out graph domains with different class cardinalities and semantics.

Target domainMethod1-shot3-shot5-shot
PubMedOurs44.154.158.9
FedAvg39.242.644.8
ogbn-ArxivOurs14.520.624.8
FedAvg13.114.818.7
WikiCSOurs33.850.952.2
FedAvg26.533.1344.9

Thank you for the rebuttal. I appreciate the authors' detailed clarifications and additional experimental results. The analysis of domain calibration and comparisons with centralized training help clarify the method's effectiveness. The additional inference-time measurements, explanations of baseline adaptation under non-aligned label spaces, and leave-one-domain-out evaluation further address my concerns regarding efficiency, fairness, and generalization. Overall, I find the responses convincing and believe they have substantially strengthened the empirical evaluation. I am satisfied with the authors' responses and I will keep the positive score.

Thank you for confirming your concerns are resolved. We sincerely appreciate your constructive comments and insightful suggestions and will incorporate them into our revision

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

This paper proposes FedDARC, a method that enables transferable prediction from a few labeled nodes without globally aligned class semantics or a fixed-output classifier.

To achieve this, they first extend graph FDG to heterogeneous source and target graphs with different class cardinalities and semantics and design domain characterization, letting the same reference encoder update gradients virtually on client data and using the optimization responses to represent features of different graph domains. Then, they use residual calibration to adjust node representations based on domain state, making the same encoder more adaptive to the graph's current domain. After that, they propose a task-relative relational inference method, which compares query nodes with all labeled context nodes to obtain their relations, then sums the attention for the same class to estimate that class's probability.

The authors compared the results on multiple target graph datasets and supported domain characterization and calibration through ablation.

Critical strengths

This paper could be a good algorithmic contribution because it has interesting and useful novelty in problem settings with non-aligned label spaces, which is a real problem in federated graph learning.

Critical weaknesses

None

Questions for authors

Some clarifications are needed regarding the experiments on the baseline adaptation protocol:

  • How do the baselines adapt to different classifier output dimensions under non-aligned label spaces?
  • What parameters are shared?
  • How is the target graph adapted in the few-shot setting?
6 comments

We thank the Reviewer for the constructive feedback and helpful suggestions.


Q1 Baseline Adaptation

  • Non-Aligned Label Spaces. Existing FL/FDG baselines typically rely on fixed-output classifiers, which cannot be directly shared across domains with different class cardinalities and semantics. Our evaluation therefore considers their adapted versions rather than assuming globally aligned class identities.

  • Adaptation Protocol. As described in Appendix C.1 and C.5\color{#c05a18}\text{Appendix C.1 and C.5}, baselines requiring target-specific adaptation are further trained using only the few labeled target context nodes to accommodate the target label space. All methods use the same source-domain training data, target context–query splits, and label budgets, with approximately equal total federated communication budgets. Target query labels are never used for training or adaptation.

These protocols allow the baselines to make predictions in the target label space without requiring semantic correspondence with source-domain classes.


Q2 Shared parameters.

Clients share the graph encoder parameters θ\theta, representation-calibration parameters ψ\psi, and relational-classification projections WQW_Q and WKW_K, which are aggregated during federated training.

The domain-state projection PϕP_{\phi} is learned separately during initialization and subsequently frozen. Raw graph data, labels, context–query partitions, and task-relative label mappings remain local and are not shared.


Q3 Few-shot target adaptation.

At inference, the few labeled target context nodes serve two purposes:

-Domain-State Estimation: Their optimization responses, computed using the fixed reference encoder, are used to derive the target domain state, which guides representation calibration.

-Relational Classification: Their task-relative labels and calibrated representations provide class evidence for query-to-context matching and prediction.

Importantly, FedDARC keeps all learned model parameters fixed during target inference. The temporary reference updates are discarded, no target-specific classifier is trained, and target query labels are used only for evaluation.

Thank you for the response. The answers on FedDARC’s shared parameters and its inference procedure are clear.

But my questions are about the baselines, and I still have two details I'd like to ask about.

  1. During federated training on the source graphs, the four clients have 10, 3, 40, and 5 classes. How is the fixed-output head of each baseline (e.g., FedAvg, FGOOD) handled across these clients?
  2. On the target graph, is a new head initialized for the target class count, and is only the head fine-tuned, or is the whole model fine-tuned?

Thank you for your question. During source-domain federated training, only the shared encoder is uploaded and aggregated; classification heads remain local. On the target graph, we initialize a new head matching the target class count, freeze the source-trained encoder, and train only the new head using the labeled target context nodes.

1. Rationale for local classification heads during source-domain training.

  • (1) Choice of conventional classifiers. To assess the advantage of our classifier design, we retain conventional classification heads consistent with each baseline’s original design, rather than replacing them with FedDARC’s relational classifier. Each client’s head matches its own label space.

  • (2) Full versus encoder-only aggregation. We also compared aggregating both the encoder and classification head with aggregating only the encoder. Full aggregation substantially reduced FedAvg’s 1-shot accuracy. We therefore report the better-performing encoder-only aggregation configuration.

Source-domain aggregation strategyCora, 1-shot (%)ogbn-Products, 1-shot (%)
Encoder + classification head15.25.6
Encoder only; local classification heads30.09.8

2. Rationale for head-only target adaptation. We tested fine-tuning the entire model, but observed convergence difficulties for FedAvg in the 1-shot setting with very limited labeled data. We therefore adopt head-only adaptation for these baselines.

We will explicitly document both the source-domain aggregation and target-domain adaptation protocols in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.