acceptodds
Under review as a conference paper at ICLR 2027

Graph Domain Adaptation Does Not End with Representation Learning

Abstract

Graph domain adaptation (GDA) transfers knowledge from a labeled source graph to an unlabeled target graph under shifts in both node attributes and graph structure. Existing methods primarily adapt graph representations through propagation redesign, distribution alignment, or source-to-target transition modeling, but still rely on a single graph-propagating path for target prediction. This leaves open whether an adapted graph representation exhausts the predictive evidence available in the target domain, since the graph-aware expert and graph-free local expert may exhibit different failure modes under topological shifts. To address this limitation, we propose EviGDA, an Evidence-Augmented Graph Domain Adaptation framework that complements graph representation adaptation with a graph-free local expert. The graph-aware expert performs message passing and entropy-aware marginal alignment, while the graph-free local expert learns solely from source node features and labels without graph propagation or target alignment. The two experts are optimized independently and combined only at inference through a task-level constant probability mixture, preserving complementary evidence without joint training, learned routing, or target pseudo-labels. Extensive experiments on ten datasets and 16 transfer tasks show that EviGDA outperforms state-of-the-art baselines. Complementary prediction paths add value beyond graph representation alignment.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Domain Adaptation, Graph Neural Networks, Distribution Shift, Ensemble Learning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.25 over 4 reviews, above 96% of the others.
Reviewer_TpamOpenReview ↗
Rating
3(+2)
Confidence
2(−2)
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

In this paper, the authors study GDA from the perspective of prediction paths. They argue that existing GDA methods mainly focus on adapting graph representations, but still rely on a single graph-propagating path for target prediction. To address this issue, they propose EviGDA. EviGDA consists of a graph-aware expert built on A2GNN with entropy-aware exact-sampling alignment, which assigns higher sampling probabilities to lower-entropy target nodes during MMD alignment, and a graph-free local expert implemented as an MLP trained only on source node features and labels. The two experts are optimized independently and combined at inference through a task-level probability mixture. Experiments on multiple GDA benchmarks verify the performance of the proposed EviGDA.

Critical strengths
  • The paper considers GDA from the perspective of prediction paths, and examines whether the adapted graph representation exhausts the predictive evidence available in the target domain.

  • Experiments on multiple GDA benchmarks verify the performance of the proposed EviGDA.

Critical weaknesses
  • The paper does not propose a new technique or new theory for GDA. It mainly assembles existing GDA techniques [1,2,3]. Moreover, both parts of Proposition 1 directly follow from the chain rule of conditional entropy and the classical ambiguity decomposition, which provide limited new theoretical insight. So, the novelty of this paper is limited.
  • The claimed insight that graph domain adaptation does not end with representation learning is not well supported. The claims in the main manuscript can be largely explained by the known behavior of GNNs under heterophily [4].
  • The design of the local expert is inconsistent with the motivation of GDA. The local expert is trained only on source features and labels without any adaptation, which is essentially a source-only model.
  • Typos and presentation issues:
    • "Most closely to our concern" in Section 2;
    • the abbreviation "EAM" is used in Section 2 before it is defined;
    • "Graph 2" in Section 4.3 is not defined;
    • Figure 1 uses "Feature-Only Predictor" and θf\theta_f, which are inconsistent with "graph-free local expert" and θℓ\theta_\ell in the text.

[1] Long M, Cao Z, Wang J, et al. Conditional adversarial domain adaptation[J]. Advances in neural information processing systems, 2018, 31.

[2] Zeng H, Lyu H, Hu D, et al. Mixture of weak and strong experts on graphs[C]//International Conference on Learning Representations. 2024, 2024: 24462-24502.

[3] Lim D, Hohne F, Li X, et al. Large scale learning on non-homophilous graphs: New benchmarks and strong simple methods[J]. Advances in neural information processing systems, 2021, 34: 20887-20902.

[4] Zhu J, Yan Y, Zhao L, et al. Beyond homophily in graph neural networks: Current limitations and effective designs[J]. Advances in neural information processing systems, 2020, 33: 7793-7804.

Questions for authors

Please see weaknesses.

2 comments
AuthorsOpenReview ↗

Official Rebuttal

Dear Reviewer Tpam,

We thank the reviewer for the careful assessment.

W1: Contribution and novelty

EviGDA addresses a specific question in GDA: whether a feature-only prediction path remains useful after cross-domain graph adaptation. Our contribution is a prediction-path formulation that keeps the adapted graph expert and the feature-only expert independently optimized, then combines their predictions through a task-level probability mixture. This separation enables controlled tests of complementarity after adaptation through parameter-matched models, graph–graph ensembles, and frozen-representation readouts.

CDAN [1] conditions adversarial alignment on classifier predictions, and Mowst [2] learns node-wise cooperation between an MLP and a GNN. EviGDA preserves independent optimization and uses a mixture coefficient shared by all nodes within a transfer task. Prior work on non-homophilous graphs [3,4] motivates the value of feature-only prediction; our study examines its complementary role after source–target graph adaptation. Within the graph branch, entropy-aware exact-sampling alignment (EAM) changes target-node sampling probabilities while retaining the multi-kernel MMD estimator and nonzero sampling support for every target node.

Proposition 1 applies standard entropy and ensemble identities to distinguish the existence of predictive information beyond the adapted Graph state from the conditions under which a trained local expert improves the mixture. This distinction motivates two complementary tests: the conditional-readout comparison assesses the predictive utility of Local inputs with Graph inputs fixed, while the expert and fusion controls assess whether the trained predictors turn complementarity into improved predictions.

The local expert has a distinct role within the adaptation system. The graph branch performs cross-domain adaptation, while the local branch supplies a separately learned, source-supervised feature-only prediction path. Its value is assessed through its contribution to the combined target predictor. The comparison with the better single expert directly tests whether combining the two paths adds predictive benefit.

W2: Evidence beyond heterophily

We test complementarity after adaptation using topology interventions, matched prediction controls, and an exploratory frozen-representation readout comparison. Degree-preserving topology interventions modify target connectivity while keeping node features and local predictions fixed. At maximum corruption, Graph-only Macro-F1 decreases by 27.020 percentage points, while local-exclusive correctness increases by 14.132 percentage points, averaged over four tasks relative to their clean graphs. Across 20 task–perturbation combinations (4 tasks × 5 nonzero perturbation levels), each averaged over five seeds, the two changes have a Spearman correlation of 0.973. This correlation is descriptive because the 20 points include repeated measurements within tasks.

In the paper's reported clean-graph control experiments across 16 transfer tasks, Graph+Local improves average Macro-F1 over the parameter-matched graph control, Graph+Graph, and the better single expert by 5.130, 4.837, and 1.982 percentage points, respectively. These matched comparisons support the predictive value of combining heterogeneous paths. The supplementary width comparison in Table 2 (reported in our response to Reviewer jMhR) separately evaluates three local capacities using a common frozen Graph checkpoint.

In the paired conditional-readout comparison (Table 3, reported in our response to Reviewer AmFm), adding the Local representation and predictions to the frozen Graph inputs increased average Macro-F1 by 2.100 percentage points, with a positive task-averaged gain in each of the five runs. CtoA was effectively tied; decreases occurred on B2E, E2B, and EN2DE. Holding the adapted Graph inputs fixed, this comparison directly measures the additional predictive utility of Local inputs for the tested readout.

W3: Presentation

“Graph 2” denotes the second, independently initialized graph expert in the Graph+Graph control. We will make this definition explicit, define EAM at its first occurrence, replace “Most closely to our concern” with “Most closely related to our work,” and align Figure 1's label and notation with “graph-free local expert.”

References

[1] Long et al. Conditional Adversarial Domain Adaptation. NeurIPS, 2018. [2] Zeng et al. Mixture of Weak and Strong Experts on Graphs. ICLR, 2024. [3] Lim et al. Large Scale Learning on Non-Homophilous Graphs: New Benchmarks and Strong Simple Methods. NeurIPS, 2021. [4] Zhu et al. Beyond Homophily in Graph Neural Networks: Current Limitations and Effective Designs. NeurIPS, 2020.

Best,

Authors

Thanks for your replies. I changed my score.

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

This paper proposes EviGDA, which combines a graph domain adaptation model with an independently trained feature-only MLP. The graph expert uses message passing and cross-domain alignment, while the MLP learns from source features and labels alone. Their predictions are combined after training using a task-level probability mixture. The paper also introduces entropy-aware target sampling for alignment and analyzes expert complementarity through conditional mutual information and Brier-risk decomposition.

Critical strengths
  1. The paper provides useful evidence for the complementary value of feature-only predictions. Parameter-matched graph models, graph–graph ensembles, and degree-preserving topology perturbations help distinguish this effect from additional capacity or generic ensembling.

  2. The empirical evaluation is broad and includes uncertainty estimates. EviGDA achieves the highest Macro-F1 on 15 of 16 tasks and improves over the better individual expert by 1.98 percentage points on average. Results include five-run means, standard deviations, and paired bootstrap confidence intervals.

Critical weaknesses
  1. The current experiments do not directly isolate the effect of keeping the local input unpropagated. A comparison within the same local architecture would clarify this design choice.

  2. The local hidden width is fixed in the reported experiments. The sensitivity of fusion gains to this capacity choice remains unclear.

Questions for authors

See weaknesses

2 comments
AuthorsOpenReview ↗

Official Rebuttal

Dear Reviewer jMhR,

We thank the reviewer for the constructive comments.

W1: Local propagation

We compared local predictors with the same architecture using either the supplied features X or their one-step propagated counterparts SX. The two settings matched parameter count, training settings, graph-expert training trajectory, and fusion coefficient. Retaining X improved average fusion Macro-F1 by 2.759 percentage points, with higher five-run means on 15 of the 16 tasks. The only decrease was on E2B (−0.819 percentage points). Table 1 reports the absolute scores and paired differences. These results support preserving the supplied features in a separate local path without additional message passing.

Table 1. Task-level Macro-F1 for local predictors using X or SX. Scores are reported on a percentage scale; differences are in percentage points.

TaskXSX
AtoC81.93080.703
AtoD76.18074.086
CtoA77.34076.810
CtoD78.08076.325
DtoA75.48074.664
DtoC81.24080.087
B2E59.76058.527
E2B75.75076.569
B2U56.90050.146
E2U51.41047.231
U2B81.20074.449
U2E58.76056.724
B1toB250.88045.077
B2toB153.01046.029
DE2EN60.58057.172
EN2DE64.22063.977
Average67.67064.911

W2: Local capacity

With local hidden widths of 32, 64, and 128, fusion improved over Graph by 3.995, 4.280, and 4.225 percentage points, respectively. Fusion also improved over the better single expert by 1.459, 1.667, and 1.674 percentage points on average, respectively. Across the evaluated configurations, average fusion Macro-F1 varied by at most 0.285 percentage points. For each task and seed, the same frozen graph-expert checkpoint was used across all three widths. These results support stable average fusion gains across the tested local capacities, while task-level responses vary by transfer. Table 2 reports the corresponding absolute Macro-F1 values.

Table 2. Supplementary fixed-Graph capacity comparison: task-level Macro-F1 for local hidden widths of 32, 64, and 128. All widths share the same frozen Graph checkpoint for each task and seed. Scores are on a percentage scale; averages and differences are calculated from unrounded scores.

TaskWidth 32Width 64Width 128
AtoC82.03782.11382.055
AtoD76.30476.31776.311
CtoA77.42277.42877.465
CtoD78.17078.19478.270
DtoA75.55375.57975.550
DtoC81.29181.29281.357
B2E59.81059.85359.810
E2B75.67475.85675.856
B2U55.68456.54555.728
E2U48.73048.78548.800
U2B77.09577.97078.133
U2E57.92358.07558.268
B1toB250.84951.19351.804
B2toB152.74553.56653.361
DE2EN60.12660.64159.980
EN2DE63.72064.27864.072
Average67.07167.35567.301

Best,

Authors

Thanks for the response. My concerns have been solved.

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

EviGDA pairs a graph-aware expert, trained with entropy-guided MMD alignment, with a separately trained graph-free MLP. It mixes their predictions using one coefficient per task and ranks first on 15 of 16 transfers.

Critical strengths

The controls are well designed: a parameter-matched graph model, a Graph+Graph ensemble, and a matched uniform sampler. They show clearly that the gain comes from the MLP's different input. The method is simple but effective.

Critical weaknesses
  1. It's unclear how hyperparameters and the fusion coefficient are chosen without target labels.

  2. Some important details, such as the fusion coefficient grid and the search spaces, appear only in the appendix. Some figures, such as Figures 2 and 5, have small fonts. A few visualizations, such as t-SNE plots of the target embeddings, would also help show what each expert captures.

Questions for authors
  1. Could you briefly explain how the fusion coefficient is chosen for each task?
  2. Do you expect the method to carry over to larger graphs or multi-source settings?
2 comments
AuthorsOpenReview ↗

Official Rebuttal

Dear Reviewer sZoG,

We thank the reviewer for the helpful suggestions.

W1: Selection protocol

The specified protocol selects hyperparameters and α on a dedicated search seed using target-domain Macro-best performance. With these settings fixed, we report the mean and sample standard deviation over five consecutive evaluation runs, each assessed at its own test Macro-best checkpoint. The reproduced baselines were evaluated using the same selection and reporting protocol, within their method-specific search spaces. Target labels are used for model selection, while parameter optimization uses source-label supervision and unlabeled target information. The supplementary comparisons reuse evaluation outcomes for selection and are reported as exploratory analyses.

W2: Larger and multi-source settings

The graph branch uses sparse propagation, and entropy-aware exact-sampling alignment (EAM) caps the number of nodes used in each MMD estimate at m = min(1000, N_s, N_t). For larger graphs, the proposed extension pairs this capped alignment estimator with mini-batch or neighborhood-sampled propagation.

For multiple sources, we propose training a graph–local expert pair for each source and combining their class-probability predictions through source-level fusion. The reported experiments cover single-source, full-batch GDA.

W3: Search-space details and presentation

We will summarize the fusion-coefficient grid and training search spaces in the main text, with complete details retained in the appendix. We have enlarged the fonts in Figures 2 and 5. t-SNE visualizations offer a qualitative view of expert representations. To quantify their predictive content, we added the paired readout comparison in Table 3 (reported in our response to Reviewer AmFm), which holds the adapted Graph inputs fixed and measures the effect of including Local inputs.

Best,

Authors

Dear Authors,

Thank you for the clear clarifications. Given the well-designed controls and consistent gains, I support acceptance and am raising my rating from 3 to 4. Please incorporate the clarifications and revisions into the final version.

Best regards, Reviewer sZoG

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

This paper studies whether a graph-free local expert can further improve target prediction after graph representation adaptation. EviGDA independently trains a graph-aware expert using graph propagation and distribution alignment, and a graph-free local expert using only source node features and labels. Their predictions are combined at inference through a task-level constant probability mixture. The paper also introduces entropy-aware exact-sampling alignment and examines expert complementarity through theoretical analysis and experiments on ten datasets and 16 transfer tasks.

Critical strengths

The simple independent dual-expert framework complements existing graph domain adaptation and achieves good overall performance across multiple transfer tasks. The paper characterizes conditions under which the local expert improves the mixture risk. Parameter-matched graph models, graph–graph ensembles, and controlled topology corruption provide useful evidence of predictive complementarity.

Critical weaknesses
  1. The interpretation of the contribution is not sufficiently validated. The experiments support the predictive complementarity of the graph-free local expert, but do not sufficiently establish that these gains come from evidence beyond the adapted graph representation. Parameter-matched graph models, graph–graph ensembles, and controlled topology corruption provide useful evidence, but do not directly verify this stronger interpretation.

  2. Model selection without target labels is insufficiently specified. The paper reports selecting checkpoints according to Macro-F1 and searching over training configurations and the mixture coefficient α, but does not adequately explain the data splits and selection procedures. A labeled source validation set can be used for model selection, and unsupervised model selection criteria such as SND (Saito et al., ICCV 2021) may also serve as references. The authors should clarify how these choices are made without target labels to make the experimental protocol clear and reproducible.

Questions for authors

How are the training configurations, checkpoint, and mixture coefficient α selected without target labels? Is a labeled source validation set used, or are other model selection criteria adopted?

5 comments
AuthorsOpenReview ↗

Official Rebuttal

Dear Reviewer AmFm,

We thank the reviewer for identifying these important issues.

W1: Evidence beyond the adapted Graph representation

We froze both experts and trained paired MLP-64 readouts with inputs [ZG,pG,HL,pL][Z_G, p_G, H_L, p_L] and [ZG,pG,0,0][Z_G, p_G, \mathbf{0}, \mathbf{0}]. The zero vectors have the same dimensions as HLH_L and pLp_L, preserving the input dimension and parameter count. Within each pair, both readouts use the same initialization, source-label split, and optimization settings. The readouts are fitted using source labels. Their output probabilities are combined with the same frozen Graph predictions using the same task-level coefficient in both arms: p=(1−α)pG+αpreadoutp=(1-\alpha)p_G+\alpha p_{\mathrm{readout}}.

Adding Local inputs increased average Macro-F1 from 64.443 to 66.543, a gain of 2.100 percentage points. The task-averaged gain was positive in each of the five runs. CtoA was effectively tied; decreases occurred on B2E, E2B, and EN2DE. Holding the adapted Graph inputs fixed, this comparison directly measures the additional predictive utility of Local inputs for the tested readout. Table 3 reports the absolute scores of the resulting predictors.

Table 3. Task-level Macro-F1 for the paired readout predictors with Local or zero-Local inputs. Scores are five-run means on a percentage scale; differences are in percentage points.

TaskGraph+LocalGraph+zero-Local
AtoC81.88781.663
AtoD76.15175.960
CtoA77.34377.343
CtoD78.02077.810
DtoA75.47575.451
DtoC81.25081.063
B2E59.54459.721
E2B72.65672.729
B2U53.80244.815
E2U49.56142.984
U2B73.60767.641
U2E57.67555.430
B1toB251.62348.444
B2toB153.22947.815
DE2EN59.77259.097
EN2DE63.09463.122
Average66.54364.443

Averages and differences are computed from unrounded scores. CtoA is effectively tied at the displayed precision.

W2: Model selection

The specified protocol selects hyperparameters and α on a dedicated search seed using target-domain Macro-best performance. With these settings fixed, we report the mean and sample standard deviation over five consecutive evaluation runs, each assessed at its own test Macro-best checkpoint. The reproduced baselines were evaluated using the same selection and reporting protocol, within their method-specific search spaces. Target labels are used for model selection, while parameter optimization uses source-label supervision and unlabeled target information. The supplementary comparisons reuse evaluation outcomes for selection and are reported as exploratory analyses.

Best,

Authors

Thank you for your response. It addressed most of my questions.

My remaining concern is that the model selection protocol may affect how we interpret EviGDA’s gains. The concern is that the same target test labels are used both to select the hyperparameters, fusion weight α\alpha, and checkpoint and to report the final scores. Even when all methods follow the same protocol, these scores are best interpreted as oracle results, not as independent test results.

For example, in Table 5 of the SND paper, ADVENT outperformed AdaptSeg by 3.2 percentage points in mIoU when target labels were used for model selection. The gap was only 0.7 points when target labels were reserved for final evaluation.

The current results may therefore be insufficient to tell how much of EviGDA’s advantage would remain if model selection did not use target labels, which is an important question for practical use.

AuthorsOpenReview ↗

Model Selection Without Target Labels on Six Citation Transfers

Dear Reviewer AmFm,

Thank you for raising this point. We agree that the original scores have an oracle interpretation because target labels informed the selection of hyperparameters, the fusion weight, and checkpoints. To measure how much of EviGDA’s advantage persists when target labels are reserved for final evaluation, we completed a follow-up comparison with ADAlign, TDSS-RW, and DGSDA on all six Citation transfers.

For each transfer, we used one dedicated search seed to select each method’s hyperparameters by maximizing Soft Neighborhood Density (SND) on unlabeled target predictions. For EviGDA, we used the same criterion to select the task-specific fusion weight α. We fixed the training budget at 150 epochs in advance and used the final checkpoint for every method. We then froze the selected settings and retrained all four methods on the same five additional seeds. Target labels were used only to score the final predictions.

Table 1. Target Macro-F1 on the six Citation transfers under SND selection (mean ± standard deviation across five evaluation seeds). The final row weights the six transfers equally.

TransferADAlignTDSS-RWDGSDAEviGDA
A→C77.35 ± 1.0477.52 ± 0.5779.67 ± 0.3780.92 ± 0.13
A→D72.33 ± 0.6971.55 ± 0.2673.12 ± 0.2774.27 ± 0.53
C→A74.76 ± 0.2374.84 ± 0.4774.93 ± 0.5375.17 ± 0.15
C→D75.03 ± 0.4471.23 ± 0.2175.65 ± 0.5276.68 ± 0.21
D→A69.88 ± 0.2373.56 ± 0.3172.35 ± 0.5473.79 ± 0.24
D→C73.67 ± 0.1776.30 ± 0.6280.71 ± 0.1580.47 ± 0.09
Six-task mean73.84 ± 0.2174.17 ± 0.2976.07 ± 0.1076.88 ± 0.13

Against DGSDA, the strongest of the three baselines on average, EviGDA’s margin narrowed from 1.21 percentage points in the original comparison to 0.81 points under SND selection. Thus, about two-thirds of the original margin remained on these six transfers. EviGDA led DGSDA on five directions, while DGSDA led by 0.24 points on D→C. The average margins over ADAlign and TDSS-RW were 3.04 and 2.71 points, respectively, compared with 2.78 and 2.81 points in the original comparison.

We are sharing the completed Citation results now while experiments on the remaining transfer tasks continue. We aim to finish those runs and share the results here before the discussion closes. We will incorporate the new results and a discussion of how model selection affects the reported gains into the paper, and we will upload a revised manuscript during the discussion period if time permits.

Best, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.