{"id":"6440d0fb-521c-4d48-b916-4f72191e9bbd","arxiv_id":"2505.20089","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"Graph domain adaptation fails more when source and target graphs have different local homophily profiles, and the proposed HGDA filters and aligns homophily, heterophily, and attribute signals to improve cross-graph node classification.","lead":"This paper shows that mismatched homophily, the tendency of linked nodes to share labels, hurts graph domain adaptation and proposes a filter-based method, HGDA, that aligns three graph-signal views across domains. A reader outside graph learning should care because many real-world label-transfer problems on networks could benefit if homophily alignment is genuinely effective.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof uses a reversed Wasserstein inequality in Proposition 1: average pairwise distance is replaced by W1, which is always smaller, so the claimed upper bound does not follow as written.","rationale":"The paper's central claim is that homophily discrepancies degrade GDA performance from both empirical and theoretical perspectives, and that aligning homophilic, heterophilic, and attribute signals (HGDA) mitigates this. The empirical evidence is suggestive, but the advertised theoretical support is Theorem 1, whose proof must be valid for the claim to stand. The reader flagged Corollary 4's unstated conditional-independence assumption; I found an earlier and more basic defect in the same proof chain. Proposition 1's replacement of an average pairwise distance by the 1-Wasserstein distance reverses the inequality: W_1 is an infimum, so it is never larger than the product-coupling average used in the double sum, and it can be much smaller. The concrete two-point example makes the failure immediate and checkable. This is not a stylistic objection or a disagreement with prevailing opinion; it is an internal inconsistency in the proof. The overclaim about theoretical justification therefore cannot be accepted as written. I also note the method specification has unresolved gaps (LD is referenced but never defined, Eq. (14) omits it, and Table 4 accuracy entries disagree with Tables 1-2), which independently weaken the empirical claim. For these reasons, the reader's REJECT verdict is appropriate, and my analysis does not change it.","tokens_in":24817,"tokens_out":9579,"duration_ms":107077,"concrete_test":"Take P^F_S = P^F_T = 0.5δ_0 + 0.5δ_1 on the real line. Then W_1(P^F_S, P^F_T) = 0 because the distributions are identical, while the uniform product-coupling average used in Eq. (26) equals 0.5. Hence Eq. (27) would assert 0.5 ≤ 0, a concrete false statement. The check is to ask the authors to re-derive Proposition 1 with the correct inequality direction; if no valid derivation of Eq. (21)/(39) can be supplied, Theorem 1 is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing flaw is in Proposition 1 (Appendix A), which supplies the step turning the PAC-Bayes domain-discrepancy term into the Wasserstein/KL terms used in Theorem 1. Eq. (26) states that the double sum over node pairs approximates N_S N_T W_1(P^F_S, P^F_T). This is not a valid approximation: W_1 is the infimum over couplings, and the double sum corresponds to the uniform product coupling, which is one feasible coupling, so the double-sum average is always at least W_1. The very next display, Eq. (27), asserts that the average pairwise distance is at most W_1, i.e. the inequality runs in exactly the wrong direction. The subsequent chain in Eqs. (31)-(39) inherits this reversed direction, so D^{γ/2}_{S,T}(P;λ) is not bounded by the Wasserstein terms as claimed. Since Theorem 1's entire bound depends on Proposition 1, the paper's central theoretical claim that homophily discrepancies degrade GDA performance is not established. This is an internal mathematical inconsistency, not a disagreement with an external consensus, and it cannot be repaired by changing a constant; it requires a different derivation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses unsupervised graph domain adaptation (GDA) under node homophily discrepancy. It presents empirical evidence that subgroups with differing homophily ratios show a negative correlation between homophily distribution divergence and target classification accuracy. The authors propose HGDA, which applies homophilic (A-tilde), full-pass (I), and heterophilic (L-tilde) filters to source and target node features, aligns the resulting embeddings via three KL divergence terms, and trains with a source classifier loss plus target entropy loss. The central theoretical claim is Theorem 1, a PAC-Bayes upper bound on target margin loss in terms of homophilic signal shift, attribute shift, heterophilic signal shift, and graph-level node heterophily distribution shift. Experiments compare HGDA and three ablations against GDA baselines on Airport, ACM, Citation, Blog, Twitch, and MAG.","tokens_in":25159,"tokens_out":6314,"duration_ms":61846,"significance":"If Theorem 1 and the empirical results were sound, the paper would provide a useful conceptual contribution: explicit separation and alignment of homophilic, attribute, and heterophilic signals for GDA, with a simple and computationally light architecture. The paper also ships extensive benchmark comparisons and an efficiency analysis. However, the proof of the core bound contains a reversed inequality involving the 1-Wasserstein distance, and an appendix table contradicts the main accuracy tables; the theoretical justification therefore fails as written, and the empirical claims are not internally consistent.","major_comments":[{"comment":"The proof of Proposition 1 uses the uniform product coupling to relate the pairwise feature distance sum to W_1(P^F_S, P^F_T). Since W_1 is the infimum over all couplings and the product coupling is one feasible coupling, the average pairwise distance is always at least W_1. The text's assertion that the sum can be approximated by N_S N_T W_1, followed by Eq. (27) '<=' N_S W_1, has the inequality in the wrong direction. Consequently the chain (31)-(39) does not establish the claimed bound on D^{gamma/2}_{S,T}(P;lambda), and Theorem 1, which relies on this proposition, is unproven. This is a load-bearing error: the paper's main theoretical conclusion that homophily discrepancies degrade GDA performance is not supported.","section":"Appendix A, Proposition 1, Eqs. (26)-(27)"},{"comment":"The proof assumes conditional independence between adjacency-transformed and Laplacian-transformed features. Both are deterministic functions of the same node features X and graph structure, so conditional independence does not hold. Moreover, the chain-rule step Eq. (43) conditions on X_S in the first factor and X_T in the second, so the decomposition cannot yield the unconditional KL sum in Eq. (41). Since Corollary 4 is the bridge from feature-distribution KL to the three signal KL terms in Theorem 1, this step also fails.","section":"Appendix A, Corollary 4, Eqs. (40)-(44)"},{"comment":"In the proof of Theorem 5, the bound DKL(P_H_S||P_H_T) <= DKL(A_S X_S||A_T X_T) + DKL(X_S||X_T) + DKL(L_S X_S||L_T X_T) is asserted without derivation and is not a consequence of the previous arguments: P_H is the distribution of label-derived node heterophily values, not a feature-signal distribution. Additionally, the final display Eq. (54) omits the heterophily KL term, whereas the theorem statements (Theorems 1 and 5) include it; the proof is internally inconsistent about which terms the bound contains.","section":"Appendix A, Eq. (52)"},{"comment":"The efficiency experiment reports HGDA A->C accuracy 0.326, while Table 2 reports HGDA A->C accuracy 0.829; other entries disagree similarly (e.g., PA A->D accuracy 0.562 in Table 4 versus 0.752 in Table 2). The table therefore cannot be read as a reliable report of the same experiments. Furthermore, no standard deviations are given anywhere, and the per-task hyperparameter choices in Table 5 are not accompanied by a validation protocol, leaving the possibility of selection on the test set unaddressed. These issues undermine confidence in the empirical claims.","section":"Appendix D, Table 4"}],"minor_comments":[{"comment":"The correspondence between the three KL alignment terms and the theoretical terms is mislabeled: Z_H is introduced in Section 4.1 as the output of the heterophilic filter and Z_F as the full-pass filter, but Section 4.2 says KL(Z_S_H||Z_T_H) aligns 'graph attributes' (DKL(X_S||X_T)) and KL(Z_S_F||Z_T_F) aligns 'heterophilic signal' (DKL(L_S X_S||L_T X_T)); these labels are swapped relative to the definitions.","section":"Section 4.2"},{"comment":"The overall loss is given as L = L_H + alpha L_S + beta L_T, but the surrounding text mentions combining L_H, L_S, 'LD' and L_T; L_D is never defined.","section":"Section 4.3, Eq. (14)"},{"comment":"Section 4 heading 'Methology' should be 'Methodology'; 'According to Theorem 7' in Section 4.2 should refer to Theorem 1; Algorithm 1 line 1 has an incomplete sentence ('by running .'); and Figure 5 caption lists 'HGDAL' twice, presumably with the fourth variant intended as HGDAH.","section":"General typos"},{"comment":"The table is titled 'Airport datset' but contains Citation rows (A->D, A->C, C->D), indicating that the efficiency experiments were not cleanly separated by dataset in the presentation.","section":"Appendix D, Table 4"}],"recommendation":"reject","confidential_remarks":"The core of the theoretical bound (Proposition 1, first inequality) is explicitly adapted from the authors' own ICLR 2025 paper (Fang et al., 2025) and related work; the genuinely new heterophily term is introduced via an unproved assertion (Eq. 52). Given the reversed-inequality error, the theoretical novelty cannot be assessed. The empirical contradictions in Table 4 suggest the manuscript was not carefully checked before submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read this with the stress-test note in hand. My verdict: the central theoretical claim does not hold as written, and the experimental reporting has internal inconsistencies, but the empirical observation about local homophily distribution shift is new and plausible.\n\nWhat's new: nobody in the cited GDA literature studies local homophily distribution shift or proposes aligning homophilic, heterophilic, and attribute signals separately. The empirical analysis in Figures 1 and 2 is a reasonable motivation, and the three-filter architecture is a simple, sensible design. The gains on MAG (8.5 points over PA) are large and consistent across the ten transfers, which suggests the method is doing something real—though with the caveat below.\n\nThe proof of Theorem 1 is not sound. Proposition 1 (Eq. 26) assumes the double sum over node pairs approximates N_S N_T W1. That is wrong: W1 is an infimum over couplings, and the uniform product coupling is just one feasible coupling, so the average pairwise distance is always at least W1. The very next display runs the inequality in the opposite direction. So the bound on D^{γ/2}_{S,T} does not follow. Theorem 1 inherits this error. The additional step bounding D_KL(P^H_S || P^H_T) by the same three signal KL terms (Eq. 52) is asserted without derivation. Together these mean the paper's headline theoretical claim—homophily discrepancies degrade GDA performance—is not established by this proof. This is an internal math problem, not a taste disagreement, and it cannot be fixed by changing a constant.\n\nTwo smaller issues: the method section says the overall loss is L = L_H + α L_S + β L_T, but L_D is referenced and never defined. And Appendix D Table 4 gives HGDA accuracy 0.326 for A→C (and C→D) while Table 2 reports 0.829 and 0.779. No standard deviations are reported anywhere despite the appendix saying five runs.\n\nThe empirical story might survive a corrected proof—the correlation plot in Figure 2 is the kind of evidence that could stand alone—but as submitted, the load-bearing theoretical claim is unsupported and the tables are inconsistent. The method specification gap is minor by comparison, but real.\n\nWho this is for: someone working on GDA who cares about heterophily-aware alignment. It deserves a serious referee because the problem is well-motivated, the idea is new within the subfield, and the empirical results are promising enough to warrant a careful revision.\n\nRecommendation: send to peer review, but expect major revision—the proof needs to be corrected or removed/re-scoped as empirical motivation, L_D needs to be defined, and the appendix tables need to be reconciled. I would not desk-reject, but I also would not trust the current Theorem 1 or the efficiency table.","headline":"The empirical motivation is fresh and the three-view alignment is a reasonable design, but Theorem 1's proof has a reversed Wasserstein inequality that collapses the theoretical claim, and the appendix tables contradict the main results.","tokens_in":25617,"tokens_out":2443,"would_cite":false,"duration_ms":25541,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that mismatched homophily between source and target graphs is a hidden driver of graph domain adaptation failure, and that aligning homophilic, heterophilic, and attribute signals separately restores performance.","keywords":["graph domain adaptation","homophily","heterophily","graph neural networks","cross-network node classification","distribution shift","graph signal filtering"],"falsifier":"Pick a benchmark pair such as U to B, estimate each KL term empirically, and check whether $D_{\\mathrm{KL}}(P_F^S \\parallel P_F^T) \\leq D_{\\mathrm{KL}}(A_S X_S \\parallel A_T X_T) + D_{\\mathrm{KL}}(X_S \\parallel X_T) + D_{\\mathrm{KL}}(L_S X_S \\parallel L_T X_T)$ holds, and separately whether the asserted bound on $D_{\\mathrm{KL}}(P_H^S \\parallel P_H^T)$ holds; failure of either inequality would show the stated proof chain does not go through.","tokens_in":24646,"feed_emoji":"🕸️","tokens_out":7690,"duration_ms":75007,"temperature":0.7,"pith_summary":"This paper sets out to show that graph homophily, the tendency of connected nodes to share labels, is a decisive and overlooked factor in graph domain adaptation (GDA), where a classifier is learned on a labeled source graph and applied to an unlabeled target graph. It documents that source and target graphs in standard benchmarks differ more sharply within homophilic and heterophilic subgroups than in their aggregate homophily, and that target accuracy falls as subgroup homophily divergence rises. It then proves a probabilistic generalization bound stating that target loss is controlled by four separate shifts: homophilic signal, attribute signal, heterophilic signal, and the graph's heterophily distribution. Accordingly it proposes HGDA, which applies homophilic, full-pass, and heterophilic filters to each graph and aligns the corresponding channels with KL-divergence losses. If the claim holds, GDA methods should stop aligning a single blended representation and instead match these three graph-signal channels separately, and HGDA should keep its reported accuracy gains across Airport, ACM, Citation, Blog, Twitch, and MAG.","feed_headline":"Separate homophily alignment lifts graph domain adaptation","feed_subtitle":"Source and target graphs diverge inside homophilic and heterophilic groups, and aligning the three channels fixes it.","key_machinery":"The central object is the four-term KL decomposition of feature-distribution shift: $D_{\\mathrm{KL}}(P_F^S \\parallel P_F^T) \\leq D_{\\mathrm{KL}}(A_S X_S \\parallel A_T X_T) + D_{\\mathrm{KL}}(X_S \\parallel X_T) + D_{\\mathrm{KL}}(L_S X_S \\parallel L_T X_T)$, together with the graph-level heterophily distribution shift $D_{\\mathrm{KL}}(P_H^S \\parallel P_H^T)$. Carried by Theorem 1, this bound says a deterministic classifier's target loss is controlled by source margin loss, a complexity term, and the sum of the square roots of those KL terms. HGDA operationalizes the decomposition with a homophilic filter (the normalized adjacency $\\tilde{A}$), a full-pass filter (the identity $I$), and a heterophilic filter (the normalized Laplacian $\\tilde{L}$), aligning the resulting embeddings channel-wise with three KL losses.","core_discovery":"The paper's central claim is that GDA performance is governed by four independent shift terms: homophilic signal, attribute signal, heterophilic signal, and graph-level node heterophily distribution. Theorem 1 upper-bounds target classification loss by source margin loss plus a penalty proportional to the square roots of the KL divergences of these four terms. To exploit the bound, HGDA applies three filters, the normalized adjacency, the identity, and the normalized Laplacian, to the source and target graphs and adds pairwise KL alignment losses between the corresponding channels. Empirically, homophily subgroup ratios negatively correlate with target accuracy, and HGDA achieves the best average accuracy across Airport, ACM, Citation, Blog, Twitch, and MAG benchmarks.","pith_inferences":["The subgroup homophily divergence could be measured before training and used as a cheap diagnostic of whether standard GDA is likely to fail on a given source-target pair.","The conditional-independence assumption behind the KL decomposition is testable; if it fails, a corrected bound would need an extra cross-term between adjacency- and Laplacian-filtered features, which would not necessarily invalidate the method's empirical gains.","Because the three filters decompose the graph signal into low-, full-, and high-frequency channels, the same channel-wise alignment idea could be carried over to arbitrary spectral filter banks, including learnable or dataset-adaptive filters.","Matching the graph-level heterophily distribution directly, for example by reweighting target nodes or constraining label entropy conditional on degree, could complement the filter alignment and further tighten the bound."],"forward_implications":["Methods that align only a single blended representation will keep failing in exactly the node subgroups whose homophily ratios diverge most between source and target graphs.","The four-term bound predicts that reducing any one of the four shifts, homophilic, attribute, heterophilic, or heterophily distribution, reduces the upper bound on target loss, so channel-wise alignment is the natural objective.","Because the heterophily-distribution term is a fixed graph statistic, two graphs can look globally similar yet still require adaptation; benchmark pairs such as ACM3 to ACM4 are cases in point.","The consistent gains of the three-filter combination over each single-filter variant imply that homophilic and heterophilic information are complementary and both need explicit alignment."],"supporting_citations":[{"why":"Supplies the probabilistic adaptation bound that Theorem 1 builds on as its starting inequality.","marker":"[Ma et al., 2021]"},{"why":"Provides the concentration inequality that turns Wasserstein distances into the KL-divergence terms appearing in the bound.","marker":"[Bobkov & Götze, 1999]"},{"why":"Supplies the expected loss discrepancy bound used in the first inequality of Proposition 1.","marker":"[Fang et al., 2025]"},{"why":"Justifies treating the normalized adjacency as a low-pass filter and the Laplacian as a high-pass filter for the method's three-channel design.","marker":"[Nt & Maehara, 2019]"},{"why":"UDAGCN is the dual graph convolutional baseline whose adversarial alignment HGDA is compared against on the benchmarks.","marker":"[Wu et al., 2020]"},{"why":"JHGDA is the hierarchical GDA baseline that defines the previous state of the art on several of the reported experiments.","marker":"[Shi et al., 2023]"},{"why":"SpecReg is the spectral-regularization GDA baseline whose theory-grounded objective HGDA extends and outperforms.","marker":"[You et al., 2022]"},{"why":"Defines the local node homophily ratio used to construct the empirical subgroup analysis and the homophily distribution plots.","marker":"[Zhu et al., 2020b]"}],"fun_headline_variants":["Homophily mismatch limits graph domain adaptation","Align homophily to bridge graph domains","Mixed filters align homophily for GDA","Graph adaptation hinges on homophily alignment","Homophily-aware domain alignment lifts accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes, without stating it in the main text, that adjacency-filtered and Laplacian-filtered features are conditionally independent given the raw features, an assumption that fails when the same features and graph structure feed both filters; it also asserts the heterophily KL bound in Eq. (52) without derivation.","fun_headline_variants_meta":{"raw":{"variants":["Homophily mismatch limits graph domain adaptation","Align homophily to bridge graph domains","Mixed filters align homophily for GDA","Graph adaptation hinges on homophily alignment","Homophily-aware domain alignment lifts accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000165,"raw_usage":{"total_tokens":1182,"prompt_tokens":812,"completion_tokens":370,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":428,"completion_tokens_details":{"reasoning_tokens":303}},"tokens_in":428,"tokens_out":370,"duration_ms":4483,"temperature":1.0,"reasoning_tokens":303,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:59:39.101533+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Pick a benchmark pair such as U to B, estimate each KL term empirically, and check whether $D_{\\mathrm{KL}}(P_F^S \\parallel P_F^T) \\leq D_{\\mathrm{KL}}(A_S X_S \\parallel A_T X_T) + D_{\\mathrm{KL}}(X_S \\parallel X_T) + D_{\\mathrm{KL}}(L_S X_S \\parallel L_T X_T)$ holds, and separately whether the asserted bound on $D_{\\mathrm{KL}}(P_H^S \\parallel P_H^T)$ holds; failure of either inequality would show the stated proof chain does not go through.","supporting_citations":[],"review_version":1}