acceptodds
Under review as a conference paper at ICLR 2027

Joint Domain Adaptation and Unknown Class Recognition on Heterophilic Graphs

Abstract

Graph domain adaptation becomes challenging when heterophily coexists with target classes absent from the labeled source graph. This setting requires learning transferable representations while avoiding the alignment of unknown target nodes with known source classes. Existing methods for heterophily or open-set adaptation focus on their respective challenges, leaving their joint treatment insufficiently explored. We propose HODA, a unified framework for eterophilic pen-set omain daptation. HODA uses multiscale spectral encoding to preserve complementary low-pass and high-pass information across propagation orders. Using the resulting representations, we adapt source class prototypes to the target distribution and introduce an unknown prototype for target private classes. Agreement between classifier predictions and prototype-based clustering selects target pseudo labels and excludes nodes jointly predicted as unknown from domain alignment. We also identify and correct inconsistent label encodings across the processed WebKB graphs, enabling semantically consistent cross domain evaluation. Experiments across multiple datasets and transfer directions demonstrate competitive performance in known-class classification and unknown-class recognition.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Domain Adaptation, Open-Set Learning, Heterophily

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_5LpkOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HODA for open-set graph domain adaptation under heterophily. The method combines multiscale low-pass and high-pass filtering with prototype adaptation and classifier–clustering consensus to classify known target nodes and reject unknown ones. Experiments evaluate the framework on Blog, WebKB, and MAG.

Critical strengths
  1. The experiments include heterophilic GNNs as well as closed-set and open-set GDA baselines. HODA improves performance on several reported transfer tasks, and the ablations help identify the effects of high-pass filtering, prototype adaptation, and consensus selection.

  2. Correcting the inconsistent class-label encodings across WebKB graphs is useful. The authors recover the original semantic labels and establish a consistent mapping across universities, addressing a concrete issue that could otherwise confound cross-domain comparisons.

Critical weaknesses
  1. Restrictive feature-space assumptions. The source and target graphs share the same feature dimension and use shared input MLP. This leaves feature heterogeneity, an important challenge in cross-domain graph learning [1], outside the evaluation. It remains unclear whether HODA can distinguish known classes affected by changes in feature semantics from genuinely unknown classes. An evaluation on different feature spaces or shared language-model encodings of raw text would strengthen its practical relevance.

  2. Limited architectural novelty. Low-pass, high-pass filtering and identity channels are already well established in heterophilic graph learning [2]. I do not see a clear architectural advance in applying these filters at multiple propagation orders and combining their outputs with MLPs. The paper needs to establish what this design contributes beyond an existing multi-pass graph encoder in heterophilic graphs combined with open-set adaptation.

  3. Insufficient evaluation on larger heterophilic graphs. WebKB contains only 183–251 nodes per graph, and Blog contains roughly 2,300–2,900. MAG adds scale, but the paper does not report homophily statistics for its evaluated splits. Larger established heterophilic benchmarks, such as arXiv-year and snap-patents [3], could support more demanding transfer evaluations. Runtime and memory measurements are also missing, leaving scalability unclear.

  4. Limited applicability of the theoretical analysis. Theorem 2 assumes a regular graph, binary classes, identical local homophily at every node, and restrictions on the interference spectrum. These conditions are quite restrictive. Real graphs can contain neighborhoods with very different homophily levels, for which different filters may be appropriate. The analysis provides limited guidance for this node-level variation or for multiclass problems. I would find an explanation of filter suitability across different neighborhoods more useful. Theorem 1's observation that additional views cannot increase the ideal joint risk is also a straightforward consequence of the nested hypothesis-class assumption.

References: [1] Liu, H., et al. “One for All: Towards Training One Graph Model for All Classification Tasks.” International Conference on Learning Representations (ICLR), 2024.

[2] Luan, S., et al. “Revisiting Heterophily for Graph Neural Networks.” Advances in Neural Information Processing Systems (NeurIPS), 2022.

[3] Lim, D., et al. “Large Scale Learning on Non-Homophilous Graphs: New Benchmarks and Strong Simple Methods.” Advances in Neural Information Processing Systems (NeurIPS), 2021.

Questions for authors
  1. How would HODA handle different source and target feature dimensions or semantics?

  2. What does the proposed encoder offer beyond existing heterophily-aware architectures?

  3. How does HODA perform on larger heterophilic graphs, and what are its runtime and memory costs?

7 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/5)

We sincerely thank you for your detailed and rigorous review. We have addressed each of them below.


1. Restrictive feature-space assumptions. (W1/Q1)

Response: Thank you for raising this point. We agree that feature heterogeneity is an important challenge in cross-domain graph learning. However, it is orthogonal to the problem studied in this paper. We follow the standard setting of graph domain adaptation, where the source and target graphs share a feature space, and all baselines and benchmarks in our evaluation adopt the same setting. Our focus is on the difficulty that arises within this setting when structural heterophily and target-private classes occur together. HODA does not address different feature dimensions or semantics by itself. In such cases, the two graphs would first need to be mapped into a common input space, after which HODA can be applied without modification.

Within a shared feature space, the concern that known classes affected by feature shift may be confused with genuinely unknown classes is the ambiguity that HODA is designed to resolve. Both kinds of nodes can receive low confidence, so HODA does not rely on confidence alone. It adapts each class prototype toward the target distribution according to its reliability and selects a node only when the classifier and the prototypes agree. We will discuss feature heterogeneity as future work in the revised version.


2. Limited architectural novelty. (W2/Q2)

Response: Thank you for this comment. We agree that low-pass, high-pass, and identity channels are well established in heterophilic graph learning, and we do not claim the encoder as a new architecture. However, these encoders were developed for a single labeled graph, and combining them with open-set adaptation is non-trivial. Under heterophily, an uncertain target prediction may come from a poorly transferred known-class node or from a genuinely unseen class, and existing methods cannot tell the two apart. Heterophilic GNNs neither align domains nor recognize unknown classes. Closed-set DA methods wrongly align unknown nodes to known classes. Open-set DA methods reject low-confidence nodes as unknown, although many known nodes also have low confidence under heterophily. The results below report ACC / HS (%) averaged over the transfer tasks of each benchmark (Tables 1, 2, 6 and 7 of the paper), with MAG averaged over both unknown-class settings.

CategoryMethodBlogWebKBMAGCitation
Heterophilic GNNsACM55.36‾\underline{55.36} / 56.25‾\underline{56.25}74.82 / 71.9644.06 / 39.3059.56 / 62.30
H2GCN53.62 / 54.8475.36‾\underline{75.36} / 72.95‾\underline{72.95}44.18 / 39.92‾\underline{39.92}60.78 / 64.19‾\underline{64.19}
Closed-set DAADAlign33.35 / 29.2946.04 / 38.0942.43 / 37.9060.79‾\underline{60.79} / 57.22
DGSDA51.88 / 52.3773.74 / 68.9641.76 / 36.2158.64 / 60.20
RSGDA40.39 / 39.5667.31 / 59.9630.36 / 14.9041.54 / 43.45
Open-set DAUAGA25.21 / 17.3753.15 / 45.4844.70 / 36.7359.80 / 58.98
GraphRTA26.87 / 21.7953.95 / 46.9346.34‾\underline{46.34} / 35.1559.74 / 61.73
OursHODA59.64 / 61.2180.28 / 76.9850.79 / 42.6266.31 / 67.72

Existing heterophily-aware encoders are the strongest baselines, but they still trail HODA by 2.70 to 4.96 points in HS, since they handle neither domain shift nor unknown classes. Closed-set DA methods are less reliable under heterophily, and even the heterophily-aware RSGDA performs poorly on MAG. Open-set DA methods perform well on the homophilic Citation benchmark but drop sharply on Blog and WebKB.

AuthorsOpenReview ↗

Official Comment by Authors (2/5)

To examine what our design contributes beyond an existing multi-pass encoder combined with open-set adaptation, we replace each part of HODA with its standard counterpart, one at a time: (1) ACM encoder replaces our multiscale low- and high-pass encoder with a tuned ACM-GCN; (2) Standard DEC replaces the reliability-weighted mixing of source and target prototypes with freely learned DEC centers; (3) MSP / Entropy / Energy replace classifier-prototype consensus with confidence-based selection. The results (ACC / HS, %) are summarized below.

VariantBlogWebKBCitation
HODA59.64 / 61.2180.28 / 76.9866.31 / 67.72
ACM encoder57.06 / 57.6570.55 / 63.8964.58 / 67.86
Standard DEC58.45 / 60.0675.99 / 63.9053.78 / 54.01
MSP58.74 / 60.1675.41 / 72.5866.39 / 63.58
Entropy58.17 / 59.6776.27 / 72.8065.47 / 60.38
Energy58.05 / 59.1374.80 / 63.8664.14 / 64.92

None of the standard counterparts is reliable across benchmarks. The ACM encoder matches ours on the homophilic Citation benchmark but falls behind under heterophily, especially on WebKB. Standard DEC is comparable on Blog but drops substantially on WebKB and Citation. Confidence-based selection yields lower HS than consensus on all three benchmarks. These results show that an existing encoder and standard open-set components cannot be combined directly in the joint setting.

To understand these failures, we examine which target nodes each variant selects on WebKB, using the quantities in Theorem 1. Here, β\beta is the fraction of known-class nodes wrongly excluded from alignment, η\eta is the fraction of unknown nodes among those retained, and Precision is the accuracy of the known-class pseudo labels. All values are averaged over five seeds and six transfer tasks.

Variantβ\betaη\etaPrecisionHS
HODA0.2390.2130.77776.98
ACM encoder0.2280.3950.54863.89
Standard DEC0.4710.1570.87063.90
MSP0.2690.2640.70272.58
Entropy0.2790.2420.72072.80
Energy0.4720.1730.83063.86

With the ACM encoder, far more unknown nodes are retained and the pseudo labels are much less accurate, which indicates that known and unknown nodes are poorly separated in its representations. Unlike ACM, our encoder uses parameter-free filters at multiple propagation orders and keeps the channels separate. The other variants fail for different reasons. Standard DEC and energy-based selection exclude nearly half of the known-class nodes, while MSP and entropy selection exclude more known nodes and retain more unknown ones than consensus. These results show that combining an existing heterophily-aware encoder with open-set adaptation is not sufficient. Our encoder does not introduce a new filter. Its contribution is to provide representations on which prototype adaptation and consensus selection remain reliable under heterophily. We will add these results to the revised version and clarify our contributions accordingly.

AuthorsOpenReview ↗

Official Comment by Authors (3/5)

3. Insufficient evaluation on larger heterophilic graphs. (W3/Q3)

Response: Thank you for these suggestions. We address the three points in turn.

Homophily. The homophily, the class counts after the split, and the share of unknown nodes are listed below for all graphs. Homophily is computed on the original classes.

BenchmarkGraphEdge homophilyNode homophilyNodes per known classNodes per unknown classShare of unknown nodes
WebKBCornell0.1230.11138 / 16 / 3082 / 1754.1%
Texas0.0610.05733 / 1 / 30101 / 1865.0%
Wisconsin0.1780.15570 / 10 / 32118 / 2155.4%
BlogBlog10.3990.389399 / 320 / 419 / 392377 / 39333.5%
Blog20.4000.390486 / 457 / 502 / 466515 / 47034.0%
CitationACMv90.8000.7941916 / 2767 / 2110803 / 176427.4%
Citationv10.8600.8322262 / 2323 / 2004690 / 165626.3%
DBLPv70.8200.8061188 / 1808 / 1307332 / 84921.5%

For MAG, the numbers of known and unknown nodes under the two open-set settings are as follows.

GraphEdge homophilyNode homophilyKnown / unknown nodes (1 unknown class)Share of unknown nodesKnown / unknown nodes (5 unknown classes)Share of unknown nodes
US0.5160.54970,709 / 61,84946.7%65,654 / 66,90450.5%
CN0.5300.58738,021 / 63,93162.7%31,628 / 70,32469.0%
DE0.5290.55321,614 / 21,41849.8%19,567 / 23,46554.5%
JP0.5630.59514,957 / 22,54160.1%13,329 / 24,16964.5%
RU0.7680.8034,874 / 27,95985.2%4,288 / 28,54586.9%
FR0.5690.57313,917 / 15,34552.4%12,886 / 16,37656.0%

arXiv-year and snap-patents. These are valuable benchmarks for heterophilic learning on a single graph. However, they are not directly suitable for domain adaptation, which requires separate source and target graphs that share the same classes. Our benchmarks instead follow those commonly used in graph domain adaptation under heterophily and in open-set graph domain adaptation: Blog and Citation [1, 2], WebKB [2, 3], and MAG [1, 4]. To our knowledge, none of these works evaluates on arXiv-year or snap-patents. Among these benchmarks, MAG provides the scale, where the largest graph has more than 130,000 nodes and nearly 700,000 edges.

Runtime and memory. We first analyze the complexity of HODA. The multiscale filters require 2J2J sparse matrix multiplications per graph, which cost O(J∣E∣d)O(J \vert \mathcal{E} \vert d) and are computed once before training. During training, the encoder applies MLPs to the precomputed views without further message passing, so each epoch costs O((ns+nt)(2J+1)dh)O((n_s + n_t)(2J+1)dh), where hh is the hidden dimension. Prototype computation and the DEC loss cost O(nt(C+1)h)O(n_t (C+1) h), and the alignment loss is computed on at most 512 sampled nodes per domain, independent of the graph size. The per-epoch cost of HODA is therefore linear in the number of nodes, and the dependence on the number of edges appears only in the one-time preprocessing.

We also measure the training time and the peak GPU memory of all methods on one NVIDIA L4 GPU with their reported configurations. The results on B1$\rightarrow$B2 and on the largest MAG task, US$\rightarrow$CN, are listed below.

CategoryMethodB1$\rightarrow$B2 time (s)B1$\rightarrow$B2 memory (GB)US$\rightarrow$CN time (s)US$\rightarrow$CN memory (GB)
Heterophilic GNNsACM2.00.038.90.3
CMGNN1.40.181051.3
HeroFilter1.80.051870.9
H2GCN10.50.77939.8
HIMP3.70.12792.7
Closed-set DAADAlign4.60.62226.2
DiffGDA28812.662313.5
Pair-Align960.24117313.4
DGSDA1234.43647.7
DFT791.17898.7
RSGDA1.90.06191.2
Open-set DAUAGA18.60.087982.0
GraphRTA9.91.1431.5
OursHODA3.50.12686.0

H2GCN is implemented in TensorFlow, and its memory is reported by the TensorFlow allocator. HODA is 3 to 17 times faster than the closed-set and open-set DA methods on US$\rightarrow$CN, except RSGDA and GraphRTA, and its memory is comparable to or lower than that of the alignment-based methods. Its cost is of the same order as that of the heterophilic GNNs, which perform no adaptation, and it is faster than H2GCN, the strongest baseline. The multiscale filters take less than 0.1 seconds on every task. Across all 34 tasks, HODA trains in under 90 seconds and uses at most 6.5 GB of memory. We will add the complexity analysis and these measurements to the revised version.

Reviewer_26AxOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies open-set graph domain adaptation on heterophilic graphs. The model has to classify target nodes from the source classes and mark nodes from new target classes as unknown. The method, HODA, has three parts. It first builds identity, low-pass, and high-pass views of the node features up to propagation order J and encodes each view with its own MLP. It then moves DEC-style source class prototypes toward the target domain, with a reliability score r_c that decides how much weight the target estimate gets. An extra unknown prototype is built from the target nodes with the highest predicted unknown probability. Finally, nodes on which the classifier and the prototypes agree become pseudo-labels, and nodes that both mark as unknown are left out of an MMD alignment loss. Before adaptation, a warm-up stage masks the true-class logit of each source node and trains the model to predict the unknown class instead. Experiments on Blog, WebKB, MAG and Citation compare HODA with 13 baselines from heterophilic GNNs, closed-set GDA, and open-set GDA, using ACC, HS and AUC over five seeds.

Critical strengths

1.The results are consistent across datasets. HODA has the best mean HS on all 18 tasks in Tables 1 and 2 and the best mean ACC on 17 of them. The paper itself points out the one exception, CN→RU. I recomputed the average gains on page 8 from the tables and they are correct. On Blog, HODA beats ACM by 4.28 ACC and 4.96 HS, and on WebKB it beats H2GCN by 4.91 ACC and 4.03 HS. Some gains are large, such as 61.94 against 53.69 HS over DGSDA on B1→B2, and HODA's HS standard deviations in Table 2 are all 1.40 or lower. The 13 baselines include many recent methods. This breadth is the main point in the paper's favor.

2.The WebKB label fix is useful beyond this paper. Table 5 shows that the processed Geom-GCN files give the same label ID to different classes across universities. Label 2, for example, means faculty in Cornell, project in Texas and student in Wisconsin. Without a fix, transfer between these graphs matches different classes, and treating the last two labels as unknown removes different classes in each graph. Other work on cross-domain WebKB can reuse this mapping. The code is also shared through an anonymous link.

Critical weaknesses
  1. The experiments do not show that the specific design of HODA explains the gains, as opposed to its encoder and its explicit unknown-class training. Table 3 shows that removing prototype adaptation or consensus filtering hurts. The rest of the evidence is weaker. On several tasks, HODA is barely ahead of heterophilic GNNs that do no adaptation at all. It leads ACM by 0.28 ACC and 0.44 HS on B2→B1, CMGNN by 0.55 ACC and 0.50 HS on WI→TX, and H2GCN by 0.61 HS on CO→TX. Each of these gaps is smaller than HODA's own standard deviation. Only AUC on B2→B1 shows a clear gain as 77.22 against 71.41.

  2. Theorem 1 assumes that the 0-1 disagreement function lies in the RKHS of the kernel with norm at most B_J. On a connected space, a 0-1 function that is not constant cannot lie in the RKHS of a continuous kernel such as the Gaussian kernel in Eq. 11. On a finite set of distinct nodes, it can, but B_J may then be very large, and the paper never estimates it. Eq. 11 averages three per-channel squared MMD terms, which matches an additive kernel on the joint representation, and the paper should state this. The result \gamma_J ≤ \gamma_j only concerns the best possible joint risk.

  3. The tuning protocol is not described, which matters a lot when target labels are not available. The paper lists search ranges for more than ten hyperparameters, including J, six loss weights, and the unknown candidate ratio. It does not say how values were chosen, whether target labels were used, or which values were picked for each task. The values of T_w, T_a, \delat, and \gamma in Algorithm 1 are also missing. Figure 4 shows that ACC is sensitive to several of these, and the best J changes from task to task.

Questions for authors

Q1. How were hyperparameters and checkpoints chosen for HODA on each task, and were target labels used at any point, including early stopping, the choice of J, and the rejection thresholds? Please list the chosen values for each task along with T_w, T_a, \delta, and \gamma, and explain how the baseline settings and the MSP or entropy thresholds were chosen.

Q2. Which kernel and hypothesis class make the RKHS assumption in Theorem 1 hold with a reasonable B_J? Is the kernel behind Eq. 11 the additive kernel over channels?

Q3. For the WebKB fix, how many nodes were matched to the original pages, and how were ties and unmatched nodes handled? Do all baselines use the corrected labels and the same splits, and how much do results change after the fix? Please also report the homophily, the class counts after the split, and the share of unknown nodes for each graph.

5 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/3)

We sincerely thank you for your thorough review and insightful suggestions. We respond to each point below.


1. The experiments do not show that the specific design of HODA explains the gains. (W1)

Response: Thank you for this careful reading. We agree that on the three tasks you point out, the margin over the best heterophilic GNN is within the standard deviation of HODA, and we do not claim a significant gain on these individual tasks. Our claim concerns consistency across tasks. A baseline can be close to HODA on a particular task, but none of them remains close across benchmarks. The results (ACC / HS, %) of the three baselines you mention, averaged over the transfer tasks of each benchmark, are summarized below, together with the average rank in HS over all 34 tasks and 14 methods. MAG is averaged over both unknown-class settings.

MethodBlogWebKBMAGCitationAverage rank
ACM55.36 / 56.2574.82 / 71.9644.06 / 39.3059.56 / 62.303.74
CMGNN51.81 / 51.9274.08 / 71.3943.10 / 36.3856.15 / 58.307.97
H2GCN53.62 / 54.8475.36 / 72.9544.18 / 39.9260.78 / 64.192.68
HODA59.64 / 61.2180.28 / 76.9850.79 / 42.6266.31 / 67.721.00

ACM is close to HODA on B2$\rightarrow$B1 but trails by 3 to 5 points in HS on every benchmark on average. CMGNN is close on WI$\rightarrow$TX, which is its only task within one point of HODA, and its average rank is close to eighth. H2GCN is the most consistent baseline, yet HODA outperforms it on all 34 tasks, by 3.30 points in HS on average. The gains are therefore not tied to particular tasks.

To examine whether these gains come from the specific design rather than from the encoder and the explicit unknown-class training alone, we replace each adaptation in HODA with its standard counterpart, one at a time: (1) ACM encoder replaces our multiscale low- and high-pass encoder with a tuned ACM-GCN; (2) Standard DEC replaces the reliability-weighted mixing of source and target prototypes with freely learned DEC centers; (3) MSP / Entropy / Energy replace classifier-prototype consensus with confidence-based selection. The results (ACC / HS, %) are summarized below.

VariantBlogWebKBCitation
HODA59.64 / 61.2180.28 / 76.9866.31 / 67.72
ACM encoder57.06 / 57.6570.55 / 63.8964.58 / 67.86
Standard DEC58.45 / 60.0675.99 / 63.9053.78 / 54.01
MSP58.74 / 60.1675.41 / 72.5866.39 / 63.58
Entropy58.17 / 59.6776.27 / 72.8065.47 / 60.38
Energy58.05 / 59.1374.80 / 63.8664.14 / 64.92

Each replacement shows the role of the corresponding component. The multiscale encoder separates known and unknown nodes under heterophily, so replacing it hurts on Blog and WebKB but not on the homophilic Citation benchmark. Prototype adaptation keeps the prototypes anchored to the source class centers, whereas the freely learned centers of Standard DEC reject many known-class nodes, which leads to large drops on WebKB and Citation. Consensus selection requires the classifier and the prototypes to agree, whereas confidence alone cannot distinguish poorly transferred known nodes from unknown ones, which lowers HS on all three benchmarks. We will add the cross-task summary and these results to the revised version.

AuthorsOpenReview ↗

Official Comment by Authors (2/3)

2. The RKHS assumption in Theorem 1, the constant BJB_J, and the kernel behind Eq. (11). (W2/Q2)

Response: Thank you for these precise comments. We agree with all of them. A non-constant 0-1 function cannot lie in the RKHS of a continuous kernel on a connected space, so the RKHS assumption in Theorem 1 is an idealization. We also agree that Eq. (11) corresponds to an additive kernel on the joint representation, and that the paper should state this explicitly. To address these points, we add the following proposition. It is stated for the additive kernel behind Eq. (11) and does not require the disagreement function to lie in the RKHS.

Proposition 1. Let krk_r be the Gaussian RBF kernel on channel r∈{I,L,H}r \in \lbrace I, L, H \rbrace, and let Ps,rP_{s,r} and Qt,rQ_{t,r} be the distributions of the channel-rr representations of source nodes and of target nodes retained for alignment. Let kJ=13∑rkrk_J = \frac{1}{3}\sum_{r} k_r be the additive kernel on the joint representation space, with RKHS KJ\mathcal{K}_J. Define

Lalign⋆=13∑r∈{I,L,H}MMDkr2(Ps,r,Qt,r),\mathcal{L}_{\mathrm{align}}^{\star} = \frac{1}{3}\sum_{r \in \lbrace I,L,H \rbrace} \mathrm{MMD}^2_{k_r}(P_{s,r}, Q_{t,r}),

for B>0B > 0, ϵapp(B)=sup⁡h,h′∈HJ inf⁡∥g∥KJ≤B(EPt,k(J)∣dh,h′−g∣+EPs(J)∣dh,h′−g∣),\epsilon_{\mathrm{app}}(B) = \sup_{h,h' \in \mathcal{H}_J}\ \inf_{\Vert g \Vert_{\mathcal{K}_J} \le B} \Big( \mathbb{E}_{P_{t,k}^{(J)}}\vert d_{h,h'} - g \vert + \mathbb{E}_{P_s^{(J)}}\vert d_{h,h'} - g \vert \Big), where Pt,k(J)=Pt(Z(J)∣Y<C)P_{t,k}^{(J)} = P_t(Z^{(J)} \mid Y < C).

Assume Pr⁡t(Suc)>0\Pr_t(S_u^c) > 0 and β<1\beta < 1. Then, for every h∈HJh \in \mathcal{H}_J and every B>0B > 0,

Rtos(h)≤min⁡{(1−πu)AB(h), C(h)}+U(h),R_t^{\mathrm{os}}(h) \le \min\big\lbrace (1-\pi_u)\mathcal{A}_B(h),\ \mathcal{C}(h) \big\rbrace + \mathcal{U}(h),

where AB(h)=Rs(h)+BLalign⋆+2B(η+β)+ϵapp(B)+λJ,\mathcal{A}_B(h) = R_s(h) + B\sqrt{\mathcal{L}_{\mathrm{align}}^{\star}} + 2B(\eta+\beta) + \epsilon_{\mathrm{app}}(B) + \lambda_J, C(h)=Lk(h)+(1−πu)[ρkϵcon+1−ρk]\mathcal{C}(h) = L_k(h) + (1-\pi_u)[\rho_k\epsilon_{\mathrm{con}} + 1 - \rho_k] and U(h)=Lu(h)+πu(1−ρu)\mathcal{U}(h) = L_u(h) + \pi_u(1-\rho_u).

Proof. The kernel kJk_J is positive semidefinite with kJ(z,z)=13∑rkr(zr,zr)=1k_J(z,z) = \frac{1}{3}\sum_{r} k_r(z_r, z_r) = 1, so κJ=1\kappa_J = 1.

Since the squared MMD is linear in the kernel and krk_r depends only on channel rr,

DJ2=MMDkJ2(Ps(J),Qt(J))=13∑rMMDkr2(Ps,r,Qt,r)=Lalign⋆.\mathcal{D}_J^2 = \mathrm{MMD}^2_{k_J}\big(P_s^{(J)}, Q_t^{(J)}\big) = \frac{1}{3}\sum_{r}\mathrm{MMD}^2_{k_r}(P_{s,r}, Q_{t,r}) = \mathcal{L}_{\mathrm{align}}^{\star}.

Fix B>0B > 0 and h,h′∈HJh, h' \in \mathcal{H}_J, and write P=Pt,k(J)P = P_{t,k}^{(J)} and Q=Ps(J)Q = P_s^{(J)}. For any gg with ∥g∥KJ≤B\Vert g \Vert_{\mathcal{K}_J} \le B, the reproducing property and the Cauchy-Schwarz inequality give ∣EP[g]−EQ[g]∣≤BMMDkJ(P,Q)\vert \mathbb{E}_P[g] - \mathbb{E}_Q[g] \vert \le B \mathrm{MMD}_{k_J}(P, Q). Adding and subtracting gg,

EP[dh,h′]−EQ[dh,h′]≤EP∣dh,h′−g∣+BMMDkJ(P,Q)+EQ∣dh,h′−g∣.\mathbb{E}_P[d_{h,h'}] - \mathbb{E}_Q[d_{h,h'}] \le \mathbb{E}_P\vert d_{h,h'} - g \vert + B \mathrm{MMD}_{k_J}(P, Q) + \mathbb{E}_Q\vert d_{h,h'} - g \vert.

Taking the infimum over gg gives

EP[dh,h′]≤EQ[dh,h′]+BMMDkJ(P,Q)+ϵapp(B).\mathbb{E}_P[d_{h,h'}] \le \mathbb{E}_Q[d_{h,h'}] + B \mathrm{MMD}_{k_J}(P, Q) + \epsilon_{\mathrm{app}}(B).

The rest follows the proof of Theorem 1 with DJ=Lalign⋆\mathcal{D}_J = \sqrt{\mathcal{L}_{\mathrm{align}}^{\star}} and κJ=1\kappa_J = 1.

Proposition 1 removes the RKHS assumption of Theorem 1, which you point out cannot hold exactly for the Gaussian kernel. The disagreement function only needs to be approximated by a function in KJ\mathcal{K}_J, and the approximation error appears explicitly as ϵapp(B)\epsilon_{\mathrm{app}}(B). In addition, BB is no longer a constant to be estimated. It is a free parameter, and the bound holds for every B>0B > 0: a larger BB reduces ϵapp(B)\epsilon_{\mathrm{app}}(B) but increases the weight on the discrepancy. The bound is informative when little probability mass lies near the decision boundaries, since a smooth function with moderate norm can then approximate the disagreement function well. The kernel behind Eq. (11) is indeed the additive kernel kJk_J over channels, and the discrepancy term is exactly the channel-wise loss. We also agree that λJ≤λj\lambda_J \le \lambda_j only concerns the best achievable joint risk and does not imply that the other terms improve with JJ. We will state this explicitly and add Proposition 1 to the revised version.

AuthorsOpenReview ↗

Official Comment by Authors (3/3)

3. The tuning protocol is not described. (W3/Q1)

Response: Thank you for pointing this out. We describe the protocol below and will add it to the revised version.

Model selection. For each task, we search the hyperparameters of HODA within the stated ranges and report the configuration that achieves the best HS on the target graph. The same protocol is applied to every baseline, so all methods are compared under the same condition. This is a common practice in open-set domain adaptation, where no labeled target validation set is available.

Checkpoints. No early stopping or checkpoint selection is used. For every task and every method, we report the model at the last epoch.

Rejection thresholds. HODA has no rejection threshold. For the baselines that require unknown rejection based on MSP or predictive entropy, we search the threshold and report the best result, following the same protocol as above.

Chosen values. The hyperparameter settings, including TwT_w, TaT_a, Δ\Delta, and γ\gamma, are provided together with the code in our supplementary material, which is why they were not listed in the paper. We will make the complete code and all configurations publicly available upon acceptance.


4. Details of the datasets. (Q3)

Response: Thank you for these questions. We answer them in order.

Matching. Every node in the three graphs is matched to its original WebKB page by its feature vector. There are no unmatched nodes and no ties. The feature vectors are unique in Cornell and Texas. In Wisconsin, two groups of nodes share identical features, but the nodes in each group carry the same label, so no ambiguity arises. The correction is a relabeling of whole classes and leaves the node order, features, and edges unchanged. It changes the label IDs of 0 nodes in Cornell, 48 in Texas, and 230 in Wisconsin, following the mapping in Table 5. All baselines use the same corrected label files and the same open-set split as HODA.

Effect of the correction. Before the correction, the same label ID refers to different semantic classes in different graphs. As a result, the labels of about 26% of the target nodes are inconsistent with the source for Cornell and Texas, about 91% for Cornell and Wisconsin, and 100% for Texas and Wisconsin. The last two label IDs, which define the unknown classes, also refer to different classes in each graph: student and project in Cornell, student and faculty in Texas, and faculty and project in Wisconsin. For example, faculty is a known class in Cornell but would be counted as unknown in Texas. Under the original IDs, all methods therefore perform poorly, and their results reflect the label mismatch rather than the quality of the methods. For this reason, we report results only on the corrected labels, and a comparison before and after the correction would not be informative.

Statistics. The homophily, the class counts after the split, and the share of unknown nodes are listed below for all graphs. Homophily is computed on the original classes. Since the correction only renames classes within each graph, the homophily of each WebKB graph is the same before and after the correction.

BenchmarkGraphEdge homophilyNode homophilyNodes per known classNodes per unknown classShare of unknown nodes
WebKBCornell0.1230.11138 / 16 / 3082 / 1754.1%
Texas0.0610.05733 / 1 / 30101 / 1865.0%
Wisconsin0.1780.15570 / 10 / 32118 / 2155.4%
BlogBlog10.3990.389399 / 320 / 419 / 392377 / 39333.5%
Blog20.4000.390486 / 457 / 502 / 466515 / 47034.0%
CitationACMv90.8000.7941916 / 2767 / 2110803 / 176427.4%
Citationv10.8600.8322262 / 2323 / 2004690 / 165626.3%
DBLPv70.8200.8061188 / 1808 / 1307332 / 84921.5%

For MAG, the numbers of known and unknown nodes under the two open-set settings are as follows.

GraphEdge homophilyNode homophilyKnown / unknown nodes (1 unknown class)Share of unknown nodesKnown / unknown nodes (5 unknown classes)Share of unknown nodes
US0.5160.54970,709 / 61,84946.7%65,654 / 66,90450.5%
CN0.5300.58738,021 / 63,93162.7%31,628 / 70,32469.0%
DE0.5290.55321,614 / 21,41849.8%19,567 / 23,46554.5%
JP0.5630.59514,957 / 22,54160.1%13,329 / 24,16964.5%
RU0.7680.8034,874 / 27,95985.2%4,288 / 28,54586.9%
FR0.5690.57313,917 / 15,34552.4%12,886 / 16,37656.0%

We will add these details and the per-class counts of MAG to the revised version.


We hope the additional results and analysis address your concerns. We would be glad to discuss any remaining questions.

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

This paper investigates open-set graph domain adaptation under heterophily, where knowledge must be transferred from a labeled source graph to an unlabeled target graph containing previously unseen classes. The authors propose HODA, a unified framework integrating multiscale spectral encoding, cross-domain prototype adaptation, and classifier–clustering consensus. The method combines low-pass and high-pass graph signals across multiple propagation orders, adapts source class prototypes to the target domain, and employs prediction agreement to select pseudo-labels and exclude unknown-class candidates from domain alignment. Experiments on several graph benchmarks, including Blog, WebKB, MAG, and Citation, demonstrate competitive performance against existing heterophilic graph learning and graph domain adaptation methods.

Critical strengths
  1. The paper addresses the joint challenges of structural heterophily, cross-domain distribution shifts, and unknown-class recognition. These challenges are practically relevant, and their combination represents a meaningful research setting. The attempt to integrate heterophilic representation learning with open-set graph domain adaptation is valuable.

  2. The authors evaluate HODA on several graph datasets and transfer tasks, comparing it against methods from different methodological categories. The reported results show strong overall performance in known-class classification and unknown-class recognition. The inclusion of ablation studies and parameter sensitivity analyses provides additional empirical evidence supporting the effectiveness of the complete framework.

Critical weaknesses
  1. Methodological novelty appears limited. Its main technical components, including multiscale spectral filtering, cross-domain prototype adaptation, DEC-based clustering have substantial foundations in existing research. The proposed framework primarily integrates and adapts these established techniques, while the specific modifications appear relatively incremental and the contribution to this field is not enough.

  2. The theoretical analysis does not adequately support the actual optimization procedure. Theorem 1 establishes an open-set risk bound involving the MMD discrepancy between the source and target distributions in the joint multiscale representation space. Eq. (11) independently minimizes MMD across the identity, low-pass, and high-pass representation channels. Aligning individual channel distributions does not generally guarantee alignment of their joint distribution. Consequently, minimizing the implemented channel-wise MMD losses does not necessarily control the joint distribution discrepancy appearing in Theorem 1. The paper does not establish a rigorous connection between these two quantities, which substantially weakens the theoretical justification of the proposed alignment strategy.

Questions for authors

Please see the weakness.

5 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/3)

We sincerely thank you for your careful review and constructive comments. We respond to each point below.


1. Methodological novelty appears limited. (W1)

Response: Thank you for this comment. We agree that the main technical components of HODA build on solid foundations, and we do not claim them as new primitives. However, each was developed under assumptions that no longer hold in our setting, where heterophily, cross-graph shift, and target-private classes occur together. Transferring them to this joint setting is therefore non-trivial. Our contribution lies in identifying the specific problem that each transfer creates and in adapting the components so that they work together. The central difficulty is that an uncertain target prediction becomes ambiguous under heterophily. It may come from a poorly transferred known-class node or from a genuinely unseen class. Existing method families fail to resolve this ambiguity for different reasons. Heterophilic GNNs handle heterophily within a single graph, but they neither align domains nor recognize unknown classes. Closed-set DA methods align all target nodes with the source, so unknown nodes are wrongly aligned to known classes. Open-set DA methods reject low-confidence nodes as unknown, but under heterophily many known nodes are also predicted with low confidence. These limitations are reflected in the results below, which report ACC / HS (%) averaged over the transfer tasks of each benchmark (Tables 1, 2, 6 and 7 of the paper), with MAG averaged over both unknown-class settings.

CategoryMethodBlogWebKBMAGCitation
Heterophilic GNNsACM55.36‾\underline{55.36} / 56.25‾\underline{56.25}74.82 / 71.9644.06 / 39.3059.56 / 62.30
H2GCN53.62 / 54.8475.36‾\underline{75.36} / 72.95‾\underline{72.95}44.18 / 39.92‾\underline{39.92}60.78 / 64.19‾\underline{64.19}
Closed-set DAADAlign33.35 / 29.2946.04 / 38.0942.43 / 37.9060.79‾\underline{60.79} / 57.22
DGSDA51.88 / 52.3773.74 / 68.9641.76 / 36.2158.64 / 60.20
RSGDA40.39 / 39.5667.31 / 59.9630.36 / 14.9041.54 / 43.45
Open-set DAUAGA25.21 / 17.3753.15 / 45.4844.70 / 36.7359.80 / 58.98
GraphRTA26.87 / 21.7953.95 / 46.9346.34‾\underline{46.34} / 35.1559.74 / 61.73
OursHODA59.64 / 61.2180.28 / 76.9850.79 / 42.6266.31 / 67.72

Heterophilic GNNs are the strongest baselines, but they still trail HODA by 2.70 to 4.96 points in HS, since they handle neither domain shift nor unknown classes. Closed-set DA methods are less reliable under heterophily, and even the heterophily-aware RSGDA performs poorly on MAG. Open-set DA methods perform well on the homophilic Citation benchmark but drop sharply on Blog and WebKB.

To further show that transferring these components is non-trivial, we replace each adaptation in HODA with its standard counterpart, one at a time: (1) ACM encoder replaces our multiscale low- and high-pass encoder with a tuned ACM-GCN; (2) Standard DEC replaces the reliability-weighted mixing of source and target prototypes with freely learned DEC centers; (3) MSP / Entropy / Energy replace classifier-prototype consensus with confidence-based selection. The results (ACC / HS, %) are summarized below.

VariantBlogWebKBCitation
HODA59.64 / 61.2180.28 / 76.9866.31 / 67.72
ACM encoder57.06 / 57.6570.55 / 63.8964.58 / 67.86
Standard DEC58.45 / 60.0675.99 / 63.9053.78 / 54.01
MSP58.74 / 60.1675.41 / 72.5866.39 / 63.58
Entropy58.17 / 59.6776.27 / 72.8065.47 / 60.38
Energy58.05 / 59.1374.80 / 63.8664.14 / 64.92

None of the standard counterparts is reliable across benchmarks. The ACM encoder matches ours on the homophilic Citation benchmark but falls behind under heterophily, especially on WebKB. Standard DEC is comparable on Blog but drops substantially on WebKB and Citation. Confidence-based selection yields lower HS than consensus on all three benchmarks. These results show that the standard components do not transfer directly to the joint setting.

AuthorsOpenReview ↗

Official Comment by Authors (2/3)

To understand these failures, we examine which target nodes each variant selects on WebKB, using the quantities in Theorem 1. Here, β\beta is the fraction of known-class nodes wrongly excluded from alignment, η\eta is the fraction of unknown nodes among those retained, and Precision is the accuracy of the known-class pseudo labels. All values are taken at the final epoch and averaged over five seeds and six transfer tasks.

Variantβ\betaη\etaPrecisionHS
HODA0.2390.2130.77776.98
ACM encoder0.2280.3950.54863.89
Standard DEC0.4710.1570.87063.90
MSP0.2690.2640.70272.58
Entropy0.2790.2420.72072.80
Energy0.4720.1730.83063.86

Each variant fails on a different term of the bound. With the ACM encoder, far more unknown nodes are retained and the pseudo labels are much less accurate, which indicates that known and unknown nodes are poorly separated in its representations. Standard DEC and energy-based selection are overly conservative: their pseudo labels are precise, but nearly half of the known-class nodes are excluded. MSP and entropy selection yield both higher β\beta and higher η\eta than consensus, since confidence alone cannot distinguish low-confidence known nodes from unknown ones. This gap is concentrated on strongly heterophilic nodes.

Taken together, these results show that solving open-set domain adaptation under heterophily requires more than combining established techniques. Our contribution is to identify why each technique breaks in this joint setting and to design the adaptation that resolves it. The multiscale encoder separates known and unknown nodes under heterophily, where a standard heterophilic encoder cannot. The reliability-weighted prototypes remain accurate on the target domain without discarding known-class nodes, unlike standard DEC. Classifier-prototype consensus selects target nodes more reliably than any confidence-based criterion. As a result, HODA achieves the best HS on every transfer task across all benchmarks, whereas no existing method or standard combination is reliable throughout. We will add these results and the discussion to the revised version and clarify our contributions accordingly.

AuthorsOpenReview ↗

Official Comment by Authors (3/3)

2. The theoretical analysis does not adequately support the actual optimization procedure. (W2)

Response: Thank you for this careful observation. We agree that aligning the marginal distributions of individual channels does not, in general, guarantee alignment of their joint distribution. To establish a rigorous connection between Theorem 1 and the implemented objective, we add the following proposition, which is stated directly in terms of the channel-wise loss.

Proposition 1. Let krk_r be the Gaussian RBF kernel on channel r∈{I,L,H}r \in \lbrace I, L, H \rbrace, and let Ps,rP_{s,r} and Qt,rQ_{t,r} be the distributions of the channel-rr representations of source nodes and of target nodes retained for alignment. Let kJ=13∑rkrk_J = \frac{1}{3}\sum_{r} k_r be the additive kernel on the joint representation space, with RKHS KJ\mathcal{K}_J. Define

Lalign⋆=13∑r∈{I,L,H}MMDkr2(Ps,r,Qt,r),\mathcal{L}_{\mathrm{align}}^{\star} = \frac{1}{3}\sum_{r \in \lbrace I,L,H \rbrace} \mathrm{MMD}^2_{k_r}(P_{s,r}, Q_{t,r}),

for B>0B > 0, ϵapp(B)=sup⁡h,h′∈HJ inf⁡∥g∥KJ≤B(EPt,k(J)∣dh,h′−g∣+EPs(J)∣dh,h′−g∣),\epsilon_{\mathrm{app}}(B) = \sup_{h,h' \in \mathcal{H}_J}\ \inf_{\Vert g \Vert_{\mathcal{K}_J} \le B} \Big( \mathbb{E}_{P_{t,k}^{(J)}}\vert d_{h,h'} - g \vert + \mathbb{E}_{P_s^{(J)}}\vert d_{h,h'} - g \vert \Big), where Pt,k(J)=Pt(Z(J)∣Y<C)P_{t,k}^{(J)} = P_t(Z^{(J)} \mid Y < C).

Assume Pr⁡t(Suc)>0\Pr_t(S_u^c) > 0 and β<1\beta < 1. Then, for every h∈HJh \in \mathcal{H}_J and every B>0B > 0,

Rtos(h)≤min⁡{(1−πu)AB(h), C(h)}+U(h),R_t^{\mathrm{os}}(h) \le \min\big\lbrace (1-\pi_u)\mathcal{A}_B(h),\ \mathcal{C}(h) \big\rbrace + \mathcal{U}(h),

where AB(h)=Rs(h)+BLalign⋆+2B(η+β)+ϵapp(B)+λJ,\mathcal{A}_B(h) = R_s(h) + B\sqrt{\mathcal{L}_{\mathrm{align}}^{\star}} + 2B(\eta+\beta) + \epsilon_{\mathrm{app}}(B) + \lambda_J, C(h)=Lk(h)+(1−πu)[ρkϵcon+1−ρk]\mathcal{C}(h) = L_k(h) + (1-\pi_u)[\rho_k\epsilon_{\mathrm{con}} + 1 - \rho_k] and U(h)=Lu(h)+πu(1−ρu)\mathcal{U}(h) = L_u(h) + \pi_u(1-\rho_u).

Proof. The kernel kJk_J is positive semidefinite with kJ(z,z)=13∑rkr(zr,zr)=1k_J(z,z) = \frac{1}{3}\sum_{r} k_r(z_r, z_r) = 1, so κJ=1\kappa_J = 1.

Since the squared MMD is linear in the kernel and krk_r depends only on channel rr,

DJ2=MMDkJ2(Ps(J),Qt(J))=13∑rMMDkr2(Ps,r,Qt,r)=Lalign⋆.\mathcal{D}_J^2 = \mathrm{MMD}^2_{k_J}\big(P_s^{(J)}, Q_t^{(J)}\big) = \frac{1}{3}\sum_{r}\mathrm{MMD}^2_{k_r}(P_{s,r}, Q_{t,r}) = \mathcal{L}_{\mathrm{align}}^{\star}.

Fix B>0B > 0 and h,h′∈HJh, h' \in \mathcal{H}_J, and write P=Pt,k(J)P = P_{t,k}^{(J)} and Q=Ps(J)Q = P_s^{(J)}. For any gg with ∥g∥KJ≤B\Vert g \Vert_{\mathcal{K}_J} \le B, the reproducing property and the Cauchy-Schwarz inequality give ∣EP[g]−EQ[g]∣≤BMMDkJ(P,Q)\vert \mathbb{E}_P[g] - \mathbb{E}_Q[g] \vert \le B \mathrm{MMD}_{k_J}(P, Q). Adding and subtracting gg,

EP[dh,h′]−EQ[dh,h′]≤EP∣dh,h′−g∣+BMMDkJ(P,Q)+EQ∣dh,h′−g∣.\mathbb{E}_P[d_{h,h'}] - \mathbb{E}_Q[d_{h,h'}] \le \mathbb{E}_P\vert d_{h,h'} - g \vert + B \mathrm{MMD}_{k_J}(P, Q) + \mathbb{E}_Q\vert d_{h,h'} - g \vert.

Taking the infimum over gg gives

EP[dh,h′]≤EQ[dh,h′]+BMMDkJ(P,Q)+ϵapp(B).\mathbb{E}_P[d_{h,h'}] \le \mathbb{E}_Q[d_{h,h'}] + B \mathrm{MMD}_{k_J}(P, Q) + \epsilon_{\mathrm{app}}(B).

The rest follows the proof of Theorem 1 with DJ=Lalign⋆\mathcal{D}_J = \sqrt{\mathcal{L}_{\mathrm{align}}^{\star}} and κJ=1\kappa_J = 1.

Proposition 1 corresponds directly to the design of HODA. The discrepancy term is now exactly the channel-wise loss minimized in Eq. (11), so the bound is stated in terms of the implemented objective. The terms η\eta and β\beta are controlled by excluding the nodes jointly predicted as unknown from alignment, and C(h)\mathcal{C}(h) is controlled by the consensus pseudo labels. Each term that can be reduced during training is thus matched by a component of HODA.

The term ϵapp(B)\epsilon_{\mathrm{app}}(B) measures how well the disagreement function can be approximated by functions in KJ\mathcal{K}_J, which are additive across channels. When the predictors rely mainly on interactions between channels, the channel marginals can be aligned while the joint distributions still differ, and this term becomes large. To examine whether this case arises in practice, we replace the channel-wise loss with a joint MMD on the concatenated channels, which is sensitive to cross-channel dependence. The results (ACC / HS, %) are summarized below.

AlignmentBlogWebKBCitation
Channel-wise MMD (HODA)59.64 / 61.2180.28 / 76.9866.31 / 67.72
Joint MMD59.48 / 61.0380.12 / 76.8466.03 / 67.52

Aligning the joint distribution directly brings no improvement, and the two strategies reach nearly identical joint discrepancy. This indicates that the extreme case does not arise in our experiments and that the channel-wise loss is sufficient in practice. We will add Proposition 1 and these results to the revised version.


We hope these responses clarify the contribution of HODA and resolve the gap between Theorem 1 and the channel-wise alignment. 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.