acceptodds
Under review as a conference paper at ICLR 2027

FAVA: Frequency Alignment across Views for Multiview Graph Domain Adaptation

Abstract

Multiview graph domain adaptation (MGDA) transfers knowledge from a labeled source graph to an unlabeled target graph, where nodes are connected through multiple relation views. We observe on real-world graphs that the source and target domains differ in their frequency composition and that this shift differs from one relation view to another. Existing MGDA methods align each view representation as a whole, leaving frequency-specific shifts uncorrected. We propose Frequency-Aware View Alignment (FAVA), which aligns the frequency composition of relation views across domains. For each relation view, FAVA learns three channels that separately capture low-frequency information, high-frequency variation, and full-pass node attributes, so that alignment can act on each frequency component individually. It then aligns these channels across all source–target view pairs, weighting each pair by its estimated shift to accommodate the view-dependent frequency shift. Experimental results on real-world graph benchmarks demonstrate the effectiveness of our method.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Domain Adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_FkmVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper tackles the problem of domain adaptation between multi-view graphs sharing the same set of views. The main contribution is frequency alignment between the target and source domains, based on a loss function that aligns the variance frequency distribution of representations from a pair of source and target views. The method outperforms previous methods on two datasets.

Critical strengths
  1. The presentation of the proposed model is clear. The technical design is also sound.
  2. The proposed method outperforms prior methods consistently in evaluation results.
Critical weaknesses
  1. Fig 1a and 1c show a substantial difference in frequency distribution between the two domains. How is the proposed frequency alignment useful in such cases?
  2. The paper fails to clearly explain how the frequency is defined in Section 1. It's easier to understand as the strength of homophily.
  3. HGDA's homophily-based domain adaptation also leverages LP/HP/FP features to align the homophily level of source and target representations. The novelty of FAVA compared to HGDA should be clarified.
  4. The variance alignment uses LP/HP/FP features processed by separate encoders. Do the encoder outputs preserve their variances? Is it valid that Eq 6 captures the distribution between frequency ranges?
Questions for authors
  1. FP preserves the full frequency spectrum, including LP and HP. The ablation study can include a variant without FP (LP and HP only).
9 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/4)

We sincerely appreciate your thoughtful and detailed feedback. We address your questions and concerns below and hope that our responses clarify the points raised in your review.

W1. The Role of Frequency Alignment under Large Domain Discrepancies

Thank you for asking us to clarify the role of frequency alignment when domain discrepancies are large.

Large cross-domain differences motivate aligning view-dependent channel compositions in the learned representations used for node classification.

  • (i) Why the observed shifts motivate adaptation.

    The substantial differences in Figs. 1(a) and 1(c) indicate different graph signal patterns in the source and target domains. These differences motivate frequency alignment: reducing their impact on learned representations to facilitate knowledge transfer from the source domain to the target domain.

  • (ii) Theoretical motivation from signal-discrepancy terms.

    This motivation has a related theoretical basis. Using HGDA's notation, Theorem 1 and Eq. (7) in its Section 3.3 give the following target-risk bound under its PAC-Bayes setting and theoretical assumptions [A]:

    LT0(ϕ)≤L^Sγ(ϕ)+1λ[2(DKL(Q∣P)+1)+ln⁡1δ]+λ4NS+λρCC02πσNSNT[DKL(ASXS∣ATXT)+DKL(XS∣XT)+DKL(LSXS∣LTXT)+DKL(PSH∣PTH)]\begin{aligned} \mathcal{L}_{T}^{0}(\phi) \leq{}& \widehat{\mathcal{L}}_{S}^{\gamma}(\phi) +\frac{1}{\lambda} \left[ 2\bigl(D_{\mathrm{KL}}(Q\\|P)+1\bigr) +\ln\frac{1}{\delta} \right] +\frac{\lambda}{4N_S} \\\\ &+\frac{\lambda\rho C C_0} {\sqrt{2\pi}\sigma N_S N_T} \Bigg[ \sqrt{D_{\mathrm{KL}}(A^S X^S\\|A^T X^T)} +\sqrt{D_{\mathrm{KL}}(X^S\\|X^T)} \\\\ &\hspace{35mm} +\sqrt{D_{\mathrm{KL}}(L^S X^S\\|L^T X^T)} +\sqrt{D_{\mathrm{KL}}(P_S^H\\|P_T^H)} \Bigg] \end{aligned}

    Here, DKL(ASXS∣ATXT)D_{\mathrm{KL}}(A^S X^S\\|A^T X^T) quantifies the discrepancy in homophilic signals, DKL(XS∣XT)D_{\mathrm{KL}}(X^S\\|X^T) quantifies the discrepancy in node-attribute signals, and DKL(LSXS∣LTXT)D_{\mathrm{KL}}(L^S X^S\\|L^T X^T) quantifies the discrepancy in heterophilic signals. The separate term DKL(PSH∣PTH)D_{\mathrm{KL}}(P_S^H\\|P_T^H) quantifies the discrepancy in the graphs' node heterophily-degree distributions [A, Section 3.3, Theorem 1]. The three signal-discrepancy terms enter the target-risk bound with positive coefficients. With other terms held fixed, larger cross-domain signal discrepancies raise the target-risk upper bound, while reducing these discrepancies tightens it. Thus, even when source-domain classification error is low, substantial signal-distribution discrepancies can keep the target-risk upper bound high. This motivates characterizing cross-domain shifts in homophilic, heterophilic, and attribute signals separately and reducing the corresponding discrepancies in learned representations.

  • (iii) The alignment object in FAVA.

    FAVA further considers the dependence of these discrepancies on relation views. Fig. 1 is used to observe cross-domain shifts across views. During training, the learned LP, HP, and FP representations of each view form a normalized channel-variance vector, and FAVA reduces cross-domain discrepancies in this composition statistic through weighted alignment of source-target view pairs (Section 3.3, p. 5, lines 248–269). Thus, the substantial differences in the figure provide a direct motivation for adaptation, while composition alignment implements this motivation in the learned representations used for node classification.

AuthorsOpenReview ↗

Official Comment by Authors (2/4)

W2. The Definition of Frequency and Its Relationship to Homophily

Thank you for highlighting the need to distinguish graph frequency from label homophily.

Graph frequency describes variations of a signal over the graph, whereas label homophily describes agreement between neighboring class labels.

  • (i) Different objects described by frequency and homophily.

    Graph frequency and homophily are closely related but describe different objects. Graph frequency describes how a given signal varies relative to the graph structure and is commonly defined through the eigenvectors and eigenvalues of the graph Laplacian. Low frequencies correspond to smoother signal variations, while high frequencies correspond to more pronounced differences between neighboring nodes. Label homophily describes the consistency of class labels between adjacent nodes [B].

  • (ii) Definitions and theoretical connection in UniFilter.

    UniFilter defines the homophily ratio as [C, Definition 2.1]

    h=\frac{ \left|\left\\{\langle u,v\rangle\in\mathcal{E}: \mathbf{y}_u=\mathbf{y}_v\right\\}\right| }{|\mathcal{E}|}

    For a normalized feature signal x∈Rn\mathbf{x}\in\mathbb{R}^n, its spectral signal frequency is defined as [C, Definition 2.2]

    f(x)=x⊤Lx2,L=I−D−1/2AD−1/2f(\mathbf{x})=\frac{\mathbf{x}^{\top}\mathcal{L}\mathbf{x}}{2}, \qquad \mathcal{L}=\mathbf{I}-\mathbf{D}^{-1/2}\mathbf{A}\mathbf{D}^{-1/2}

    For a connected graph and an optimal polynomial filter for node classification under the corresponding assumptions, Theorem 3.1 states that the filtered signal's spectral frequency is proportional to 1−h1-h, where hh is the homophily ratio [C].

  • (iii) Graph properties and signal distributions in HGDA.

    Furthermore, HGDA distinguishes the graph's heterophily distribution from the distributions of different graph signals. Using HGDA's notation, Theorem 1 and Eq. (7) give the following target-risk bound under its PAC-Bayes setting and theoretical assumptions [A, Section 3.3]:

    LT0(ϕ)≤L^Sγ(ϕ)+1λ[2(DKL(Q∣P)+1)+ln⁡1δ]+λ4NS+λρCC02πσNSNT[DKL(ASXS∣ATXT)+DKL(XS∣XT)+DKL(LSXS∣LTXT)+DKL(PSH∣PTH)]\begin{aligned} \mathcal{L}_{T}^{0}(\phi) \leq{}& \widehat{\mathcal{L}}_{S}^{\gamma}(\phi) +\frac{1}{\lambda} \left[ 2\bigl(D_{\mathrm{KL}}(Q\\|P)+1\bigr) +\ln\frac{1}{\delta} \right] +\frac{\lambda}{4N_S} \\\\ &+\frac{\lambda\rho C C_0} {\sqrt{2\pi}\sigma N_S N_T} \Bigg[ \sqrt{D_{\mathrm{KL}}(A^S X^S\\|A^T X^T)} +\sqrt{D_{\mathrm{KL}}(X^S\\|X^T)} \\\\ &\hspace{35mm} +\sqrt{D_{\mathrm{KL}}(L^S X^S\\|L^T X^T)} +\sqrt{D_{\mathrm{KL}}(P_S^H\\|P_T^H)} \Bigg] \end{aligned}

    In particular, DKL(PSH∣PTH)D_{\mathrm{KL}}(P_S^H\\|P_T^H) measures the discrepancy in node heterophily-degree distributions, whereas DKL(ASXS∣ATXT)D_{\mathrm{KL}}(A^S X^S\\|A^T X^T), DKL(XS∣XT)D_{\mathrm{KL}}(X^S\\|X^T), and DKL(LSXS∣LTXT)D_{\mathrm{KL}}(L^S X^S\\|L^T X^T) measure discrepancies in different graph signals. HGDA treats the graph's heterophily distribution and its signal distributions as distinct quantities [A, Section 3.3, Theorem 1]. FAVA accordingly constructs its alignment objective from learned representations of different filtered signals. The distributions of same-label neighbor ratios in Fig. 1 are used to observe the local smoothness of label signals and its view-dependent variation across domains, providing an intuitive diagnostic for this modeling motivation.

  • (iv) Clarification in the manuscript.

    We have added the definition of graph frequency and its relationship to homophily to Section 1 (pp. 1–2, lines 52–57).

AuthorsOpenReview ↗

Official Comment by Authors (3/4)

W3. Novelty Compared with HGDA

Thank you for asking us to clarify FAVA's methodological contribution relative to HGDA.

FAVA targets view-dependent channel-composition shifts and aligns composition vectors across source-target view pairs, rather than channel-wise embedding distributions.

  • (i) Multiview problem setting.

    The differences between FAVA and HGDA center on the research problem and the alignment object.

    First, HGDA adapts one graph in each domain, whereas FAVA addresses multiview graph domain adaptation and focuses on frequency-related cross-domain shifts that vary across relation views. Different relation views of the same graph have different structures, and the shift within the same relation view can also differ across domains. FAVA therefore considers both domain discrepancies and their dependence on the relation view.

  • (ii) Different alignment objects and view-pair weighting.

    Second, HGDA separately applies KL alignment to source- and target-domain node embeddings in the homophilic, heterophilic, and full-pass channels, and sums the three alignment terms [A, Section 4.2, Eq. (11)]. FAVA normalizes the total variances of the three channels within each view into a composition vector that characterizes the relative contributions of the learned representations from different filtering branches. It then performs weighted alignment of these composition vectors over all source-target view pairs. Corresponding views provide direct relation correspondences, while non-corresponding views receive adaptive weights according to composition similarity.

  • (iii) Methodological contribution of FAVA.

    Our contribution is to identify view-dependent frequency-related domain shifts in multiview graphs and introduce FAVA, a frequency-aware view alignment framework for MGDA. FAVA constructs a normalized channel-variance composition for each relation view, bringing alignment to the frequency-composition level of individual views. It directly aligns corresponding views and adaptively weights non-corresponding views by composition similarity, addressing view-dependent frequency shifts through composition alignment across all source-target view pairs.

W4. Independent Encoders and the Statistical Meaning of Eq. (6)

Thank you for asking about the statistical meaning of the encoded channel variances.

Eq. (6) characterizes variation across nodes in learned channel representations, and Eq. (7) expresses their contributions to concatenated total variance.

  • (i) Statistical meaning of the encoded variances.

    Independent encoders adjust the numerical values and distributions of the input features. Eq. (6) computes the total variance of the encoded node representations to characterize how the current channel representations vary across nodes. Eq. (7) then normalizes the total variances of the three channels to describe each channel's relative contribution to the total variance of the concatenated representation, directly linking the composition statistic to the learned representations used for node classification.

  • (ii) Alignment on learned representations.

    Constructing statistical alignment objectives on learned representations has precedents. HGDA aligns channel embeddings obtained after trainable transformations and nonlinear activations [A, Sections 4.1-4.2], while Deep CORAL aligns the covariance of layer activations in nonlinear neural networks [D].

  • (iii) Manuscript clarification and supplementary analyses.

    We have clarified this statistical interpretation after Eq. (7) in Section 3.3 (p. 5, lines 243–247). The centered spectral analyses and channel-variance composition comparisons are provided in Appendix L (pp. 29–30, lines 1540–1545) and Appendix J (pp. 25–27, lines 1317–1349), respectively.

Reviewer_xY2zOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

FAVA addresses multiview graph domain adaptation through low-pass, high-pass, and full-pass representation channels. It represents each view’s “frequency composition” using normalized channel variances and aligns these vectors across source–target view pairs. Experiments on 12 transfer tasks from ACM and MAG show consistent improvements over the evaluated baselines.

Critical strengths
  1. FAVA achieves the best reported Accuracy and Macro-F1 across all 12 tasks. Average gains over SAGA reach 17.3/20.2 percentage points on ACM and 5.1/6.3 points on MAG.
  2. Ablations retain the same multichannel architecture when removing alignment, supporting its practical contribution. Sensitivity and efficiency analyses provide additional evidence beyond the main comparisons.
Critical weaknesses
  1. Increasing FP output scale drives all normalized composition vectors toward the FP axis, reducing alignment loss. Inversely scaling the corresponding classifier weights preserves predictions. The paper does not establish whether training avoids this behavior.
  2. The LP interpretation needs qualification. With the chosen two-step filter, gLP(λ)=(1−λ)2g_{\mathrm{LP}}(\lambda)=(1-\lambda)^2 increases for λ>1\lambda>1. Self-loops do not guarantee that these eigenvalues are absent, so this filter is not uniformly low-pass.
  3. Alignment weights vary from 2 to 160 across tasks, alongside substantial changes in other hyperparameters. The paper does not explain how these settings are selected without target labels.
  4. There are also discrepancies between the results in Table 3 and the full model results in Appendix F. For example, the accuracy for CN→DE is reported as 0.448 in Table 3 but 0.458 in Appendix F. First, the reason for these discrepancies is unclear, raising concerns about the reliability of the experimental results. Second, these discrepancies imply very small performance gains in the corresponding ablation experiments, casting doubt on the effectiveness of the individual components.
  5. All graphs come from two academic dataset families and contain exactly two views. Broader claims about multiview frequency alignment remain weakly tested, especially under controlled frequency shifts.
Questions for authors

See the weaknesses。

4 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/3)

Thank you for your careful review and valuable comments. Below are our responses to the questions and concerns raised in your review. Please let us know if further clarification would be helpful.

W1. Channel Scaling and Degeneracy of Composition Alignment

Thank you for raising the question about channel scaling during training.

Existing parameter regularization penalizes channel-encoder amplification, and Appendix J reports the observed channel-variance compositions across tasks.

  • (i) Effect of parameter regularization.

    Actual training applies L2L_2 parameter regularization with a coefficient of 10−410^{-4} to all three channels. L2L_2 regularization penalizes the squared norm of the parameters. For example, when the final-layer parameters of a channel encoder are scaled up and the corresponding classifier weights are scaled down, the encoder-side penalty grows quadratically with the scale, while the classifier-side penalty varies inversely with its square. Their combined penalty does not remain invariant. Therefore, a reparameterization that preserves the classification output is not an invariant transformation of the complete regularized training objective, and continually increasing the channel encoder parameters incurs a corresponding regularization cost [A, Section 7.1.1].

  • (ii) Channel-variance composition analysis.

    We have added the definitions of total variance and normalized channel proportions in Appendix J (Channel Variance Composition Analysis) (p. 25, lines 1317–1349). Figures 7 to 12 (pp. 26–27) provide the complete results for all 12 transfer tasks and 48 domain-view combinations. The table below summarizes mean post-encoding channel-variance proportions (%), computed before applying the fixed channel coefficients. We average the proportions equally over the source and target relation views within each task and then over tasks in each dataset family.

    Dataset familyTransfer tasksLP (%)HP (%)FP (%)
    ACM279.756.3413.91
    MAG1053.835.7140.46

    HP and FP jointly account for 20.25% of the mean variance composition on ACM and 46.17% on MAG. Compared with ACM, MAG has a 26.55-percentage-point higher mean FP share and a 25.92-percentage-point lower mean LP share. The mean HP shares are 6.34% and 5.71%, respectively. These means show contributions from all three channels, with different channel allocations across the two dataset families.

W2. Interpretation of LP and the Smoothing Effect of Two-Step Propagation

Thank you for your comment on the interpretation of the LP branch.

Across all twelve transfer tasks, two propagation steps yield lower view-averaged variation than one step in both the source and target domains.

  • (i) Comparison protocol and variation metric.

    Appendix K (Signal Smoothness after One and Two Propagation Steps) (p. 28, lines 1460–1503) examines the practical smoothing effect of the LP branch by comparing one and two propagation steps on the actual graph signals [B, C].

    For the processed signal Zv,kuZ_{v,k}^{u} defined in Appendix K, we compute the graph-shift-based variation [D, Eqs. (4) and (10)]

    =\frac{ \left\\|Z_{v,k}^{u}-\widehat A_v^u Z_{v,k}^{u}\right\\|_F^2 }{ \left\\|Z_{v,k}^{u}\right\\|_F^2 }$$ where $k\in\\{1,2\\}$. Smaller values indicate less signal variation under the propagation operator.
  • (ii) Results across the twelve transfer tasks.

    Figure 13 (p. 29, lines 1533–1536) reports the source- and target-domain results for all 12 transfer tasks. In every task, the view-averaged variation after two propagation steps is lower than after one step. The table below reports dataset-family means: we first average the two view-specific metrics within each graph, and then average equally over the two ACM graphs or the six MAG graphs.

    Dataset familyOne stepTwo steps
    ACM0.04950.0231
    MAG0.06570.0493

    The mean variation decreases from 0.0495 to 0.0231 on ACM and from 0.0657 to 0.0493 on MAG. For individual graphs, ACM1 decreases from 0.0398 to 0.0188, CN from 0.0633 to 0.0471, and US from 0.0947 to 0.0753.

  • (iii) Smoothing through repeated neighborhood aggregation.

    One propagation step performs a weighted aggregation of a node's own features and those of its direct neighbors, while two steps further combine information within the two-hop neighborhood [B, C]. These results provide experimental evidence for the practical smoothing effect of the LP branch on the actual graph signals used in this paper.

AuthorsOpenReview ↗

Official Comment by Authors (2/3)

W3. Variation in Alignment Weights and Hyperparameter Selection without Target Labels

Thank you for highlighting the need to clarify hyperparameter selection without target labels.

Appendix I.1 provides an SND-based protocol with a shared MAG candidate space and no target-label use for candidate scoring or selection.

  • (i) Model selection in unsupervised domain adaptation.

    Under the standard unsupervised GDA setting, no labeled target-domain validation set is provided, making it difficult to select hyperparameters and the number of training epochs. The same issue arises in image domain adaptation, and previous UDA literature has noted that some studies use target-domain test performance to select hyperparameters or model checkpoints [E].

  • (ii) MAG candidate search.

    To examine performance when configurations are selected without target-domain labels, we have added a Soft Neighborhood Density (SND)-based hyperparameter-selection protocol in Appendix I.1 (pp. 23–24, lines 1238–1257). For the ten MAG transfer tasks, we perform a joint grid search over the same candidate sets:

    \lambda_t\in\\{0,0.0025,0.005,0.01,0.02\\}$$
  • (iii) Training, SND scoring, and selection.

    Each task has 30 candidate configurations, each trained for 100 epochs, with the SND score computed at the end of training. Specifically, we normalize the class-probability vectors of the target nodes and compute pairwise similarities. After excluding self-pairs, we apply temperature-scaled softmax to construct soft-neighborhood distributions, compute their mean entropy over the target nodes, and select the configuration with the highest SND score [E].

    Each task independently selects its configuration using the same candidate space and selection rule. The selected configurations are reported in Appendix I.1, Table 4 (p. 24, lines 1259–1272). Ground-truth target labels are not involved in SND scoring or configuration selection.

W4. Discrepancies between the Main-Table and Appendix Results

Thank you for pointing out this inconsistency.

The discrepancy arose from different seed-aggregation protocols. We have updated the MAG ablation results using five random seeds.

  • (i) Source of the discrepancy.

    Table 3 (p. 8) reports the average results over five random seeds, as specified in Appendix A (p. 15, lines 766–767), whereas the MAG ablation results in the previous version of Appendix F were obtained using a single fixed random seed. This accounts for the differences in the full-model results between the two locations.

  • (ii) Updated results.

    We have rerun the MAG ablation experiments over five random seeds, averaged the results using the same reporting protocol as the main experimental table, and updated Figure 4 in Appendix F (MAG results, p. 20, lines 1072–1079). For each random seed, the ablation variants perform worse than the full model under the same seed. The five-seed averages show the same trend, supporting the effectiveness of the corresponding components.

AuthorsOpenReview ↗

Official Comment by Authors (3/3)

W5. Datasets and the Multiview Configuration

Thank you for raising the question about the datasets and multiview configuration.

The ACM and MAG benchmarks comprise eight graphs and twelve transfer tasks, and Appendix M adds a controlled target-feature frequency-perturbation experiment across all twelve tasks.

  • (i) Benchmark graphs and transfer tasks.

    We use the ACM and MAG academic-network dataset families, comprising eight graphs and twelve directed transfer tasks, consistent with the MGDA benchmark setting used by SAGA [G].

  • (ii) Meta-path view counts in related benchmarks.

    Existing multiview graph-learning benchmarks typically contain two or three meta-path views. For example, HAN uses two views, PAP and PSP, on ACM, and three views, APA, APCPA, and APTPA, on DBLP [H, Section 5.1 and Table 2]. InfoMGF uses three views, BUB, BSB, and BLB, on Yelp [I], while DMG uses three views, MAM, MDM, and MWM, on Freebase [J].

  • (iii) Role of the multiview configuration.

    The multiview configuration allows us to examine both frequency-composition differences across relation views and the role of composition alignment using non-corresponding views. The twelve real-world graph transfer tasks provide evaluations of these aspects across different transfer directions.

  • (iv) Controlled frequency perturbations.

    We have added a controlled frequency-perturbation experiment in Appendix M and Table 5 (p. 31, lines 1620–1673). Increasing the perturbation strength α\alpha gives higher-frequency components in the target features greater relative emphasis, while keeping source features, graph structures, and total target input energy unchanged.

    The mean ACC/Macro-F1 results for ACM and MAG are summarized below:

    Dataset familyOriginal FAVAα=0.5\alpha=0.5α=1\alpha=1α=2\alpha=2α=5\alpha=5
    ACM0.6615/0.59950.6510/0.58750.6445/0.58000.6380/0.57150.6215/0.5415
    MAG0.5331/0.44530.5292/0.44220.5252/0.43940.5140/0.43020.4919/0.4101

    These results show that FAVA remains stable under frequency perturbations. At α=0.5\alpha=0.5, mean ACC/Macro-F1 across all 12 tasks decreases by only 0.50/0.46 percentage points.

References

[A] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016.

[B] Qimai Li, Zhichao Han, and Xiao-Ming Wu. Deeper insights into graph convolutional networks for semi-supervised learning. Proceedings of the AAAI Conference on Artificial Intelligence, 32(1):3538–3545, 2018.

[C] Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. Simplifying graph convolutional networks. In Proceedings of the 36th International Conference on Machine Learning (ICML), volume 97 of Proceedings of Machine Learning Research, pp. 6861–6871, 2019.

[D] Siheng Chen, Aliaksei Sandryhaila, José M. F. Moura, and Jelena Kovačević. Signal recovery on graphs: Variation minimization. IEEE Transactions on Signal Processing, 63(17):4609–4624, 2015.

[E] Kuniaki Saito, Donghyun Kim, Piotr Teterwak, Stan Sclaroff, Trevor Darrell, and Kate Saenko. Tune it the right way: Unsupervised validation of domain adaptation via soft neighborhood density. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 9184–9193, 2021.

[F] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. In Proceedings of the 34th Conference on Uncertainty in Artificial Intelligence, volume R16 of Proceedings of Machine Learning Research, pp. 875–884. PMLR, 2018.

[G] Ruiyi Fang, Jingyu Zhao, Shuo Wang, Ruizhi Pu, Bingheng Li, Jiale Cai, Zhihao Li, Zihao Jing, Jian Zhu, Song Tang, Charles Ling, and Boyu Wang. SAGA: Structural aggregation guided alignment with dynamic view and neighborhood order selection for multiview graph domain adaptation. In International Conference on Learning Representations, pp. 95585–95607, 2026.

[H] Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S. Yu. Heterogeneous graph attention network. In The World Wide Web Conference (WWW), pp. 2022–2032, 2019.

[I] Zhixiang Shen, Shuo Wang, and Zhao Kang. Beyond redundancy: Information-aware unsupervised multiplex graph structure learning. In Advances in Neural Information Processing Systems, volume 37, pp. 31629–31658, 2024.

[J] Yujie Mo, Yajie Lei, Jialie Shen, Xiaoshuang Shi, Heng Tao Shen, and Xiaofeng Zhu. Disentangled multiplex graph representation learning. In Proceedings of the 40th International Conference on Machine Learning (ICML), volume 202 of Proceedings of Machine Learning Research, pp. 24983–25005, 2023.

AuthorsOpenReview ↗

Looking forward to hearing from you

Dear Reviewer #xY2z,

Thank you for your careful review and for raising these methodological and experimental questions. We explained how the existing regularization penalizes increases in channel encoder parameter scale, added a channel-variance composition analysis, and compared the smoothing effects of one and two propagation steps. We also documented the SND protocol for selecting MAG configurations without target labels, updated the MAG ablation results, and evaluated controlled frequency perturbations across all 12 tasks. Your suggestions helped us clarify our design choices and provide more targeted experimental evidence for the paper’s conclusions.

With the discussion period nearing its end, we welcome any further feedback and will make every effort to address any remaining concerns..

Best regards,
The Authors of Submission #30698

Reviewer_7weFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the overlooked phenomenon in multiview graph domain adaptation (MGDA) that frequency-related domain shifts vary across relation views, and proposes the FAVA framework. The authors first conduct an empirical analysis on real-world multiview graphs, finding that within the same graph, different relation views exhibit systematic differences in frequency distribution shape and mean, and that under the same view, the source-target shift pattern also varies across views. Based on this, FAVA constructs low-pass (LP), high-pass (HP), and full-pass (FP) filtered channels separately for each relation view and encodes them, uses a normalized vector of the total variances of the three channel representations to characterize the view's "frequency composition," and then performs frequency-composition alignment over all source-target view pairs: corresponding views receive a fixed weight of 1, while non-corresponding views are adaptively weighted by frequency-composition similarity. Finally, it concatenates all view and channel representations and jointly trains a shared classifier with source-domain cross-entropy, target-domain conditional entropy, and a cross-view frequency alignment loss.

Critical strengths
  1. The problem observation is empirically grounded, and the motivation chain describe well.
  2. The characterization of frequency composition is lightweight and interpretable.
  3. The experimental coverage and strength of comparisons are reasonably described well.
Critical weaknesses
  1. The LP/HP/FP filtering occurs only at the input, and the encoders are nonlinear, so the post-encoding variance ratio is not guaranteed to still correspond to "low-frequency/high-frequency" energy proportions. The paper neither proves nor empirically tests the fidelity of this mapping.
  2. For Eq. (9), all datasets in the paper happen to have M=2, and there is no discussion of how this weight systematically shrinks as M increases for M>2, diluting the non-corresponding view alignment term, nor any sensitivity analysis with respect to M.
  3. This method assumes source and target domains share the same set of relation views M with one-to-one correspondence, assigning weight 1 to v=w. But the conclusion defers partially overlapping or unmatched view sets to future work, the most realistic cross-graph scenario without any approximate correspondence or degradation strategy. The method's applicability boundary is thus exposed by its own conclusion.
  4. For Eq. (10), the alignment loss constrains only first-order statistics, not means or covariance structure.
Questions for authors
  1. This variant sets λ_a=0 while retaining all three channels and all views since it removes the alignment objective, not the frequency-aware representation.
  2. SAGA is selected as the main reference, but whether SAGA's hyperparameter search space is equivalent to FAVA's that is unclear.
  3. The empirical analysis of frequency distributions uses target-domain labels, but the boundary with the label-free training setting is not made clear.
5 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/4)

We appreciate your detailed and constructive feedback. Below, we address your concerns and questions point by point. We are happy to provide further clarification if needed.

W1. Interpretation of Post-Encoding Variance

Thank you for raising this question about the interpretation of the encoded representations.

FAVA's variance composition measures each filtering channel's relative contribution to the total variance of the learned concatenated representation.

  • (i) Graph-filtered inputs and nonlinear encoding.

    Graph frequencies describe signal variation relative to graph structure and are defined through Laplacian eigenvalues and eigenvectors, rather than fixed temporal or spatial frequency bands [A, B]. LP extracts neighborhood smoothness, HP captures node-neighborhood differences, and FP preserves node attributes [C, D].

    Following ACM and HGDA, which use learnable transformations and nonlinear activations in different channels [E, F], FAVA separately encodes filtered inputs to learn node representations for classification.

  • (ii) Frequency characteristics after encoding.

    We additionally examine the encoded representations from a graph-frequency perspective [N]. Figure 14 in Appendix L ("Spectral characteristics after nonlinear encoding"; pp. 29–30, lines 1540–1545 and 1606–1607) shows that, among the three encoded channels, LP exhibits the strongest energy concentration in the low-frequency region, whereas HP generally exhibits higher spectral energy density at higher frequencies than the other encoded channels. FP follows the spectral variation trend of the overall signal. These observations show that the branches retain distinct relative frequency characteristics corresponding to their filtering origins after nonlinear encoding.

  • (iii) Relative contributions to concatenated variance.

    Regarding the post-encoding variance ratios, we agree that nonlinear encoders do not guarantee that these ratios remain strictly proportional to the low- and high-frequency energy proportions of the input graph signals. FAVA is not intended to preserve or reconstruct the exact input spectrum. Instead, it uses separate channel encoders to obtain node representations from different filtered inputs. Specifically, the normalized variances in Eq. (7) accurately describe each channel’s relative contribution to the total variance of that concatenated representation. This characterization reflects both the channels' filtering origins and the features learned for node classification, linking composition alignment to the representations used for prediction. We have clarified this interpretation in Section 3.3, immediately after Eq. (7) (p. 5, lines 243–247).

W2. Normalization of Non-Corresponding View Weights

Thank you for asking us to clarify the normalization of non-corresponding view weights.

At a fixed average composition similarity, each source view's total weight for non-corresponding pairs does not decrease as the number of views increases.

  • (i) Pair-count normalization.

    The factor 1/[2(M−1)]1/[2(M-1)] in Eq. (9) (Section 3.3, p. 5, lines 254–263) normalizes the number of non-corresponding view pairs. For each source view, there are M−1M-1 non-corresponding target views. As the number of views increases, the weight of an individual pair decreases, but the number of participating pairs increases simultaneously. Specifically, let ovw∈[0,1]o_{vw}\in[0,1] denote the frequency-composition similarity between views vv and ww, as defined in Eq. (8). The total weight of the non-corresponding pairs associated with source view vv is

    = \frac{1}{2(M-1)}\sum_{\substack{w\in\mathcal M\\\w\ne v}}o_{vw} = \frac12\left(\frac{1}{M-1}\sum_{\substack{w\in\mathcal M\\\w\ne v}}o_{vw}\right).$$
  • (ii) Total weight and similarity-based allocation.

    Thus, this total weight equals half the average composition similarity of the non-corresponding views. When the average similarity remains unchanged, the total weight does not decrease merely because MM increases. Moreover, the non-corresponding weights adapt to composition similarity: more similar view pairs receive larger alignment weights, giving greater emphasis to more closely related non-corresponding views.

  • (iii) View counts in existing benchmarks.

    Existing multiview graph-learning benchmarks commonly contain two or three meta-path views [O, P, Q, R]. In practice, MM is therefore typically small. Together with pair-count normalization, the overall cross-view alignment weight is not severely diluted solely by an increase in the number of views.

AuthorsOpenReview ↗

Official Comment by Authors (2/4)

W3. Shared Relation Views and Non-Corresponding View Alignment

We appreciate your comments on the shared-view setting and its applicability.

FAVA follows the shared-view setting used in prior GDA studies and performs weighted alignment over all source–target view pairs.

  • (i) Shared views in existing GDA benchmarks.

    Corresponding source–target views are common in GDA, from early single-view transfer [G] to heterogeneous and multiview adaptation. MuSDAC and HGA use shared meta-paths across domains [H, I], while SAGA shares PAP/PSP views on ACM and PAP/PP views on MAG [J]. FAVA therefore follows the shared-view settings and evaluation protocols of representative GDA studies.

  • (ii) Alignment of all source–target view pairs.

    Meanwhile, FAVA does not restrict alignment to one-to-one corresponding views. It performs weighted alignment over all source–target view pairs, assigning adaptive weights to non-corresponding pairs according to their frequency-composition similarity. Therefore, even if a view has no counterpart of the same relation type in the other domain, it can still be assigned weights based on its frequency-composition similarity to each view in that domain and participate in alignment.

  • (iii) Future scope for partially overlapping view sets.

    We recognize the research value of partially overlapping or unmatched view sets. This does not diminish the broad practical applicability of the current work. We have revised the final sentence of Section 5 (Conclusion) (p. 10, lines 503–505) to clarify that future work will explore how frequency-composition similarity can guide cross-view alignment between domains with partially overlapping or unmatched relation-view sets.

W4. Statistical Information Captured by Eq. (10)

Thank you for drawing attention to the statistical information captured by the alignment objective.

Eq. (10) uses second-order central statistics, and total variance achieves the highest mean ACC and Macro-F1 in both dataset families in our comparison.

  • (i) Second-order central statistical content.

    Eq. (10) uses the normalized total-variance composition of the encoded representations from different filtering branches. Total variance is the sum of the feature-wise variances and corresponds to the trace of the covariance matrix. The objective therefore incorporates second-order central statistical information, rather than only first-order statistics.

  • (ii) Comparison with alternative statistics.

    To examine how the choice of statistic affects transfer performance, we further compared five alternatives: the second moment, mean vector, squared norm of the mean, standard deviation, and covariance norm. The table below reports mean ACC/Macro-F1, averaged over the two ACM tasks and the ten MAG tasks, respectively. The best score for each metric is shown in bold. The complete per-task results are provided in Table 3 of Appendix H (Comparison of Alignment Statistics; p. 23, lines 1201–1231).

    Dataset familyVariance (FAVA)Second momentMean vectorSquared norm of the meanStandard deviationCovariance norm
    ACM0.661/0.5990.619/0.5480.545/0.5210.575/0.5590.602/0.5670.652/0.590
    MAG0.533/0.4450.530/0.4400.527/0.4420.529/0.4410.530/0.4430.530/0.443
  • (iii) Classification results and design rationale.

    The variance-based formulation achieves the highest mean ACC and Macro-F1 in both dataset families. Compared with the second moment, total variance improves mean ACC/Macro-F1 by 4.25/5.15 percentage points on ACM and 0.27/0.51 percentage points on MAG. Replacing total variance with the second moment reduces both metrics on all 12 transfer tasks. Moreover, total variance can be computed from feature-wise statistics without explicitly constructing the full covariance matrix, offering lower computational and storage costs than full-covariance computation.

    These results support our choice of total-variance composition: aligning the relative variation across nodes in different filtering branches effectively supports target-domain node classification. Eq. (10) is designed to match this composition statistic, rather than simultaneously matching the means and full covariance structures.

AuthorsOpenReview ↗

Official Comment by Authors (3/4)

Q1. Interpretation of the Alignment Ablation

Thank you for pointing out the ambiguity in the alignment ablation.

Setting λa=0\lambda_a=0 removes the alignment objective while retaining all three filtering channels, all relation views, and their encoders.

  • (i) What the variant removes.

    The term cross-view frequency alignment in our paper refers to the alignment objective La\mathcal{L}_a defined in Eq. (10), rather than the frequency-aware representation constructed through the LP, HP, and FP channels. Therefore, setting λa=0\lambda_a=0 removes the entire cross-view frequency alignment objective while retaining all three filtering channels, all relation views, and their encoding processes.

  • (ii) Controlled backbone and manuscript clarification.

    This ablation deliberately retains the same representation-learning backbone to isolate the contribution of adding the alignment objective, rather than evaluating the effect of removing the frequency-aware representation. We have revised the first ablation item, "w/o Frequency Alignment," in Section 4.3 (Ablation Study) (p. 8, lines 413–414) to explicitly state that this variant retains all three filtering channels, all relation views, and their encoders.

Q2. Hyperparameter Search Settings

We appreciate your request for clearer hyperparameter-search settings.

Appendix I.1 specifies the MAG candidate space, training budget, SND selection rule, and task-specific selected configurations.

  • (i) Baseline configuration and MAG grid search.

    SAGA uses the parameter configuration provided in its original paper [J]. For FAVA, we have specified the grid-search settings for MAG tasks in Appendix I.1 (pp. 23–24, lines 1238–1257). The candidate sets are λa∈1,5,10,50,100,200\lambda_a\in\\{1,5,10,50,100,200\\} for the composition-alignment weight and λt∈0,0.0025,0.005,0.01,0.02\lambda_t\in\\{0,0.0025,0.005,0.01,0.02\\} for the target conditional entropy weight. The final configuration is selected according to the SND score. Table 4 (p. 24, lines 1259–1272) separately reports the final configuration adopted for each task to distinguish the candidate search space from the selected settings. Moreover, SAGA and FAVA differ in both the number and roles of their hyperparameters.

  • (ii) Coarse-grid sensitivity results.

    Appendix I.2 and Figure 6 (pp. 24–25, lines 1275–1286 and 1311–1313) report ACC differences from the within-task best across the six composition-alignment weights for all ten MAG tasks. On most datasets, ACC varies by less than 0.5 percentage points, with task-dependent, non-monotonic responses. These curves show how performance changes over the coarse grid, rather than reporting only each task's selected configuration.

Q3. Target Labels in Offline Analysis, Training, and Model Selection

Thank you for highlighting the distinction between offline analysis, training, and model selection.

Target labels support offline analysis but are not required by FAVA's training objectives or by the supplementary SND scoring and configuration-selection procedure.

  • (i) Offline diagnostics and training objectives.

    The target labels used in the frequency-distribution analysis serve to compute class-related statistics offline and illustrate differences between the two domains. These analysis metrics are neither training inputs nor optimization objectives for FAVA. FAVA uses source-domain labels for supervised training, while the target-domain composition vectors, alignment weights, and entropy loss are computed from node representations or model predictions without requiring ground-truth target labels.

  • (ii) Model selection in unsupervised domain adaptation.

    Under the standard unsupervised GDA setting, no labeled target-domain validation set is provided. The same issue arises in conventional image domain adaptation. Prior UDA studies have noted that some benchmark studies select hyperparameters or model checkpoints using performance computed with ground-truth target labels, a practice termed oracle selection [K, L].

  • (iii) Supplementary label-free SND protocol.

    To examine transfer performance when configurations are selected without target-domain labels, we have added an SND-based label-free hyperparameter selection protocol in Appendix I.1 (pp. 23–24, lines 1238–1257). After training each candidate for 100 epochs, we construct soft-neighborhood distributions from target-node predictions, use the mean neighborhood entropy as the SND score, and select the highest-scoring configuration [K]. Table 4 (p. 24, lines 1259–1272) reports the final hyperparameter configurations selected for each task.

    Under this supplementary protocol, ground-truth target labels are not involved in SND scoring or configuration selection.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

Dear Area Chair and Reviewers,

Thank you for your time and effort in reviewing our paper and for your detailed, constructive feedback. We appreciate your recognition of the following strengths:

  • Empirically grounded motivation. The problem observation is empirically grounded, and the motivation chain is clearly presented (Reviewer 7weF).
  • Lightweight and interpretable frequency-composition characterization. It provides a concise statistical summary of relative channel contributions (Reviewer 7weF).
  • Consistent classification gains. FAVA achieves the best reported Accuracy and Macro-F1 on all 12 transfer tasks (Reviewer xY2z). Its consistent improvements over prior methods are also recognized (Reviewer FkmV).
  • Ablations that isolate the alignment contribution. Retaining the same multichannel backbone supports assessment of the alignment objective's contribution, while sensitivity and efficiency analyses provide evidence beyond the main comparisons (Reviewer xY2z).
  • Clear presentation and sound technical design (Reviewer FkmV).

We have responded point by point and supplemented the manuscript with methodological clarifications and experimental analyses. The main updates are:

  • Clarifying graph frequency, composition statistics, and the distinction from HGDA. We added the definition of graph frequency and its relationship to homophily and clarified each channel's relative contribution to concatenated total variance. Our responses distinguish FAVA's cross-view composition alignment from HGDA's channel-wise embedding-distribution alignment (Reviewers 7weF and FkmV; Section 1, pp. 1–2, lines 52–57; Section 3.3, p. 5, lines 243–247).
  • Analyzing filtering, nonlinear encoding, and channel composition. We added centered post-encoding spectral analyses for all 12 tasks and channel-variance comparisons for 48 domain-view combinations, and explained the parameter-scaling cost under existing L2L_2 regularization. One- versus two-step comparisons show lower view-averaged variation after two steps in both domains of every task (Reviewers 7weF, xY2z, and FkmV; Appendices J–L, pp. 25–30).
  • Comparing alignment statistics. We evaluated five alternatives: second moment, mean vector, squared norm of the mean, standard deviation, and covariance norm. Total variance achieves the highest mean ACC and Macro-F1 on both ACM and MAG (Reviewer 7weF; Appendix H, Table 3, p. 23, lines 1201–1231).
  • Extending channel ablations and updating MAG results. We added variants removing LP, HP, or FP; w/o FP retains only LP and HP. These removals reduce mean ACC across the 12 tasks by 11.46, 8.73, and 9.10 percentage points, respectively. We updated the MAG ablation figure and clarified its five-seed aggregation in our responses (Reviewers xY2z and FkmV; Section 4.3, pp. 8–9; Appendix F, pp. 20–21).
  • Clarifying label-free hyperparameter selection and sensitivity. We provided the SND selection procedure and selected configurations. Each MAG task uses the same 30 candidates, each trained for 100 epochs; target labels are not involved in candidate scoring or selection. We also report coarse-grid ACC-difference curves for all ten MAG tasks (Reviewers 7weF and xY2z; Appendix I.1 and Table 4, pp. 23–24, lines 1238–1272; Appendix I.2 and Figure 6, pp. 24–25).

We hope these clarifications and additions assist the Area Chair and reviewers in assessing FAVA's contribution. Thank you for your consideration. We are happy to provide further clarification if needed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.