{"id":"7bb0e5b6-b65e-4ef1-8900-d3a6c301c2a5","arxiv_id":"2501.02969","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LOHA directly contrasts low-pass and high-pass graph filters as mutual negatives and adds a spectral signal trend composite feature to reunite node features, improving self-supervised node classification.","lead":"A new self-supervised framework, LOHA, trains graph neural networks by contrasting low-pass and high-pass spectral views of the same nodes and then uses a spectral signal trend feature to pull the views back together. It reports state-of-the-art node classification on nine graphs, beating fully supervised models on several heterophilic datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's entrywise bound on filtered features (max|X'|≤max(λ')B) is invalid for general spectral filters, so the stability ranking of C- over C+ that motivates L_sf is unsupported; the theoretical core of 'harmony in diversity' is weakened, though empirical results may still stand.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: the stability of the composite feature, which motivates subtraction over addition and the whole L_sf loss, rests on Theorem 1 and on the invalid inequality max|X'| ≤ max(λ')B. I agree that this inequality is not generally valid for spectral graph filters, and the defect is specific enough to matter: the concentration exponent in Eq. 7 changes, and the claimed ranking of C- versus C+ no longer follows from the stated argument. I therefore view the theoretical justification as the most load-bearing weak point. That said, the empirical results are extensive and the ablation variants are consistent, so the invalid theorem does not force rejection by itself. The paper can be repaired either by correcting the bound and checking whether the subtraction-vs-addition conclusion survives, or by explicitly presenting the stability argument as intuition rather than proof. The missing S3GCL baseline and absence of code are secondary concerns but reinforce the conditional stance. Since my recommended verdict matches the reader's CONDITIONAL judgment, I mark the verdict as UNCHANGED.","tokens_in":13120,"tokens_out":6971,"duration_ms":65483,"concrete_test":"Take the trained LOHA filters (from Eq. 9) on Cora and Cornell. Compute M = U(g_l(Λ)-g_h(Λ))U^T, B = max|X|, and check whether max_{i,j}|MX| ≤ max(λ')B holds. Independently re-derive Eq. 7 using the correct bound max_{i,j}|MX| ≤ ||M||_∞·B. If the original inequality is violated, or if ||M||_∞ for subtraction is not smaller than for addition, Theorem 1 cannot justify C-, and the theoretical motivation for L_sf should be withdrawn or revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the theoretical justification for C-(x_l,x_h) in Eqs. 6 and 13. Theorem 1 (Eq. 7) assumes that after filtering by M = U g(Λ) U^T, max_{i,j}|X'| ≤ max(λ')·B, where max(λ') is the maximum eigenvalue of M. This inequality is not valid for a general spectral filter. For any matrix M, the correct entrywise bound is max_{i,j}|X'| ≤ ||M||_∞·B (maximum absolute row sum); for a symmetric spectral filter, ||M||_∞ can exceed the largest eigenvalue by a factor that grows with degree and filter order. For example, M = I - A_norm has maximum eigenvalue 2, but row sums can exceed 2 on high-degree graphs, and higher-order polynomial filters can amplify this gap further. If the premise of Theorem 1 fails, the concentration exponent in Eq. 7 must involve ||M||_∞ (or the spectral norm with an additional sqrt(NF) factor), not max(λ') alone. Consequently, the paper's conclusion that subtraction (C-) is more stable than addition, and hence that L_sf in Eq. 13 should be built on C-, does not follow. This is the specific point on which the 'relative stability' requirement and the 'harmony in diversity' rationale rest. Without it, L_sf is an ad hoc regularizer, and the method's gains may be due to the flexible filter parameterization in Eq. 9 rather than to the proposed composite feature.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LOHA, a self-supervised graph contrastive learning framework that directly contrasts low-pass and high-pass spectral views as negative samples, and introduces a composite feature based on a 'spectral signal trend' to reunite the separated node features. The method combines a separation loss (L_l + L_h) with a reunion loss (L_sf), and experiments on nine datasets report an average accuracy of 77.1%, surpassing previous self-supervised baselines and, on heterophilic datasets, fully supervised models. The paper includes ablations supporting the contributions of the individual loss components.","tokens_in":13544,"tokens_out":7216,"duration_ms":68474,"significance":"If the empirical claims hold, LOHA would be a state-of-the-art self-supervised node classifier, particularly strong on heterophilic graphs, and the idea of using opposing spectral filters as natural negatives without augmentation is an interesting departure from prior augmentation-based contrastive methods. The paper provides extensive experiments and ablations, including a demo experiment suggesting that the gains are not solely due to filter complexity. However, the theoretical support for the composite feature is not rigorous; the key concentration bound relies on an unjustified assumption, and the choice of subtraction over addition is not demonstrated. The contribution is therefore primarily empirical as it stands.","major_comments":[{"comment":"The paper assumes that for a graph filtering operation Ug(Λ)U^T X, the entrywise maximum of the filtered feature X' is bounded by max(λ')·B, where max(λ') is the largest eigenvalue of the filter matrix. This bound is not valid for general spectral filters: the correct entrywise bound involves the induced ∞-norm of Ug(Λ)U^T, which can substantially exceed the largest eigenvalue on graphs with high degree or for higher-order polynomial filters. For example, the filter I + \\hat{A} used in Eq. (2) has largest eigenvalue 2, but its maximum absolute row sum can be larger on irregular graphs. Since Theorem 1 and the subsequent stability comparison of C- versus C+ rest on this assumption, the theoretical conclusion that subtraction is more stable than addition is unsupported.","section":"Theoretical Analysis, Eq. (7)"},{"comment":"Even if the bound in Eq. (7) were valid, the paper immediately concludes 'we subtract low-pass and high-pass views' for the composite feature. This requires that max(λ') for (g_l - g_h) is smaller than max(λ') for (g_l + g_h), but the paper does not prove this for the filters in Eq. (9). For the simplest filters in Eq. (2), (g_l + g_h) = 2I has eigenvalue 2, while (g_l - g_h) = -2Λ has largest eigenvalue 0 but largest absolute eigenvalue 2λ_max, which can exceed 2. No analytical or empirical comparison of these quantities is provided for the actual learned filters, so the design choice of C- is not justified.","section":"Theoretical Analysis, after Eq. (7)"},{"comment":"The abstract and introduction claim that the spectral signal trend is 'relatively unaffected by changing filters' and 'focuses solely on original feature differences.' However, the composite feature in Eq. (6) is defined as ||UΔg(Λ)U^T Δx_i||, which explicitly depends on the learned filter Δg(Λ). Similarly, Eq. (5) depends on the graph degree and neighbor features, not solely on original node features. The theoretical analysis therefore applies to a quantity that is not the one used in the loss L_sf, and the claimed filter-independent stability of the composite feature is contradicted by its definition.","section":"Theoretical Analysis, Eq. (6) and Abstract"},{"comment":"The loss L_sf contrasts the learned final embedding zf_i with C-(zl_i, zh_i), but the definition of C- in Eq. (13) is a norm of a filtered version of the original feature difference Δx_i, not a function of the learned embeddings zl_i and zh_i. This is not a standard InfoNCE pairing, and the paper does not explain the mechanism by which aligning zf_i with this norm improves the learned filters or representations. As written, L_sf is closer to a regularizer than a contrastive objective; the paper should either provide a theoretical account of its effect or explicitly characterize it as a regularization term.","section":"Loss Design, Eq. (13)"}],"minor_comments":[{"comment":"The sentence 'for n1 ∈ n1_i stands for the 1-hop neighbors of node i' is garbled; it should read 'where n1_i denotes the set of 1-hop neighbors of node i.'","section":"Equation (5)"},{"comment":"The denominator of L_l includes s(zl_i, zh_i), and the denominator of L_h also includes s(zl_i, zh_i); the paper should clarify why the same-node cross-view term is treated as a negative for both losses and whether this double counting is intentional.","section":"Equation (12)"},{"comment":"The sentence 'LOHA surpasses all baseline models on 7 out of 9 benchmarks' is ambiguous because on Cora and Citeseer, the self-supervised baseline PolyGCL outperforms LOHA; please specify that LOHA is best among self-supervised baselines, and note the tie on PubMed.","section":"Table 2"},{"comment":"The mention of 'SpCo (2022)' lacks a citation; either add the reference or remove the mention.","section":"Related Work"},{"comment":"The definitions of Tr1 and Tr3 are unclear: the operations 'mean' and 'std' are applied to vector arguments, and the resulting dimensions are not specified; please clarify the notation.","section":"Equations (14) and (15)"},{"comment":"The paper uses both 'Lsp' and 'Lsf' to refer to the reunion loss (e.g., 'w/o Lsp' in the variant list versus Eq. (13)); please unify the terminology.","section":"Figure 3 and text"},{"comment":"The demo experiments in Table 1 do not specify how the band-stop and band-pass filters are integrated into the PolyGCL framework; please provide implementation details or a reference to the appendix.","section":"Table 1"},{"comment":"The claim that the sliding parameter ensures the γ_i sequences 'will not intersect' should be explained more explicitly, as the monotonicity of the cosine term may not guarantee non-intersection for all parameter values.","section":"Equation (9)"}],"recommendation":"major_revision","confidential_remarks":"The empirical evaluation is extensive and the results are encouraging, but the theoretical framework is currently not rigorous enough to support the paper's central claims about 'harmony in diversity.' The missing appendix (including the proof of Theorem 1) and the questionable assumption in Eq. (7) would need to be addressed before this can be considered for publication. The paper also has several presentation issues, including a missing reference and inconsistent notation, but these are secondary."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: LOHA is a well-executed empirical paper with a genuinely new contrastive setup, but the theoretical story that motivates its composite feature doesn't survive contact with the actual spectral filter. The experiments are the contribution; the theory should be rewritten or cut.\n\nWhat's new: treating low-pass and high-pass embeddings of the same node as mutual negatives in InfoNCE, then adding a regularization term that aligns the final embedding with a composite feature built from the difference of the two filtered signals. That direct contradiction is absent in PolyGCL and S3GCL, and the sliding cosine parameterization of the filters is a sensible small trick. The results are extensive: nine datasets, consistent gains, and beating supervised baselines on heterophilic graphs is a real empirical result. The ablation replacing PolyGCL's loss with LOHA's loss is a good control and shows the loss itself carries the improvement.\n\nSoft spots: The theoretical justification for using subtraction rather than addition in the composite feature rests on Theorem 1, which assumes max_{i,j}|X'| ≤ max(λ')·B for filtered features. That inequality is not valid for general spectral filters. The correct entrywise bound uses the max row sum of the filter matrix, which can exceed the largest eigenvalue by a factor that grows with degree and filter order. Without that bound, the concentration argument for C- being more stable than C+ doesn't go through, and the 'harmony in diversity' story becomes an ad hoc regularizer. The proof is deferred to an appendix that isn't in the arXiv version, which doesn't help. There's also a symbol inconsistency (x vs z) around Eq. 13, and the S3GCL baseline is missing despite being cited and closely related. No code release, though the method is simple enough to reimplement.\n\nThis deserves a serious referee. I'd push for major revision: fix or delete Theorem 1, add S3GCL, release code. If the authors reframe L_sf as a heuristic regularizer with strong empirical support, the paper is acceptable. The reader's conditional verdict is about right; I'd go a bit stronger on the theory problem.","headline":"Strong empirical paper on direct low/high-pass contrastive learning, but the theoretical justification for the composite feature rests on an invalid spectral bound and should be revised or cut.","tokens_in":14003,"tokens_out":2071,"would_cite":true,"duration_ms":79093,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LOHA treats low-pass and high-pass spectral views as mutual negatives and reunites features with a spectral signal trend, reaching 77.1% average accuracy and beating fully supervised models on heterophilic graphs.","keywords":["graph contrastive learning","spectral graph neural networks","low-pass filters","high-pass filters","heterophilic graphs","self-supervised node classification","spectral signal trend","Chebyshev polynomial approximation"],"falsifier":"Compute the empirical variance of the composite feature $C_-(z^l_i,z^h_i)$ for fixed nodes across training epochs while the filters update; if it drifts as much as the raw low-pass or high-pass embeddings, the stability premise of Theorem 1 is contradicted and the reunion loss is not acting as the claimed fixed landmark.","tokens_in":12958,"feed_emoji":"🕸️","tokens_out":14880,"duration_ms":126663,"temperature":0.7,"pith_summary":"The paper argues that the natural opposite behaviors of low-pass and high-pass graph filters can be used directly as contrastive views in self-supervised learning, without data augmentation. The LOHA framework pushes the low-pass and high-pass embeddings of the same node apart, treating them as mutual negatives, while a composite feature built from the spectral signal trend (the degree-normalized difference between a node's features and its neighbors' features) acts as a stable landmark to keep each node identifiable. If this works as claimed, label-free graph representation learning can match or beat fully supervised methods on heterophilic graphs, where connected nodes tend to have different labels and features. The paper reports 77.1% average node-classification accuracy over nine real-world datasets, 2.8% above the runner-up, and results that surpass fully supervised baselines on all heterophilic datasets tested.","feed_headline":"Self-supervised graph learner uses low/high-pass contrast to hit 77.1%","feed_subtitle":"The method needs no labels or augmentations yet beats fully supervised models on heterophilic graphs.","key_machinery":"The load-bearing object is the spectral signal trend $T_r(x_i)=\\left(x_i/\\sqrt{d_i}-\\sum_{j\\in n_i^1}x_j/\\sqrt{d_j}\\right)^2$, a per-node feature that measures how a node's attribute vector changes against its degree-normalized neighborhood aggregate. The composite feature $C_-(z^l_i,z^h_i)=\\|U(g_l(\\Lambda)-g_h(\\Lambda))U^\\top \\Delta x_i\\|$ combines the two views by subtraction, which Theorem 1's concentration bound suggests is the more stable choice because smaller maximum eigenvalues tighten the bound; this composite feature powers the reunion loss $L_{sf}$ that keeps a node identifiable after its low- and high-pass embeddings are pushed apart. The views themselves are generated by sliding cosine-parameterized Chebyshev polynomial filters, which interpolate the filter response and enforce strictly ordered low-pass and high-pass sequences.","core_discovery":"The paper's central claim is that the natural opposition between low-pass and high-pass spectral filters is the right source of contrastive views for self-supervised graph learning, provided the model also reunites each node's split features with a composite landmark. It defines the spectral signal trend $T_r(x_i)=\\left(x_i/\\sqrt{d_i}-\\sum_{j\\in n_i^1}x_j/\\sqrt{d_j}\\right)^2$, the squared degree-normalized difference between a node's features and its one-hop neighborhood aggregate, and uses the filtered version $C_-(z^l_i,z^h_i)=\\|U(g_l(\\Lambda)-g_h(\\Lambda))U^\\top \\Delta x_i\\|$ as the landmark. The training loss $L=L_l+L_h+\\mu L_{sf}$ directly makes $z^l_i$ and $z^h_i$ mutual negatives while pulling the full-pass embedding $z^f_i$ toward the low-pass, high-pass, and composite features. The paper reports 77.1% mean accuracy across nine benchmark graphs, 2.8% above the runner-up, and LOHA outperforms fully supervised models on every heterophilic dataset tested.","pith_inferences":["A natural extension is to generalize the spectral signal trend beyond one-hop neighbors, with k-hop or attention-weighted aggregates, which may help on graphs with longer-range heterophily.","The composite-feature recipe is separable from the specific Chebyshev parameterization, so it could be grafted onto other spectral self-supervised methods that currently maximize agreement between augmented views instead of opposing low- and high-pass filters.","The theory predicts a stability ordering between subtraction and addition composites; directly comparing $C_-$ and $C_+$ across graphs with different spectral radii would test whether the concentration argument, rather than the flexible filters, drives the gains."],"forward_implications":["Self-supervised graph representation learning can drop hand-crafted data augmentations: the low-pass/high-pass opposition supplies the contrastive signal directly.","Heterophilic graphs become a favorable regime for contrastive learning: LOHA reports higher accuracy than fully supervised baselines on Cornell, Texas, Wisconsin, Actor, Chameleon, and Amazon ratings.","Ablation results indicate the three loss terms are complementary: removing the direct contrast term causes the largest drop, removing the reunion term hurts most on homophilic graphs like Cora, and the sliding filter parameterization adds consistent small gains.","The framework transfers across filter parameterizations, since substituting LOHA's loss into the runner-up PolyGCL's pipeline still improves accuracy on most datasets."],"supporting_citations":[{"why":"Supplies the PolyGCL baseline and the low-pass/high-pass filter-contrastive setup that LOHA directly compares to and ablates against.","marker":"Chen, Lei, and Wei 2024"},{"why":"Provides the Chebyshev interpolation with reparameterized coefficients used as the base polynomial approximation for the filter views.","marker":"He, Wei, and Wen 2022"},{"why":"Provides the InfoNCE-style contrastive loss formulation that the LOHA losses are built from.","marker":"Chen et al. 2020"},{"why":"Source of the distributional assumptions (identical, independent, bounded features) used in the concentration bound of Theorem 1.","marker":"Ma et al. 2022"},{"why":"Defines Dirichlet energy, whose degree-normalized neighborhood difference inspires the spectral signal trend in Eq. 5.","marker":"Karhadkar, Banerjee, and Montufar 2023"},{"why":"S3GCL is the competing low-pass/high-pass self-supervised framework whose cosine-parameterized filters Eq. 9 extends with a sliding term.","marker":"Wan et al. 2024"}],"fun_headline_variants":["Low/high-pass contrast helps LOHA hit 77.1% and beat supervised","LOHA: spectral contrastive learning beats supervised on heterophily","Self-supervised LOHA uses low/high-pass views for 77.1% accuracy","Spectral GNN LOHA: 77.1% with no labels, outdoes supervised"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The design of the reunion loss assumes that the spectral signal trend between low-pass and high-pass views remains stable enough to act as a fixed landmark while filters are learned; the paper's proof of that stability is deferred to an appendix and rests on a bound on filtered features that is not generally valid for spectral graph filtering.","fun_headline_variants_meta":{"raw":{"variants":["Low/high-pass contrast helps LOHA hit 77.1% and beat supervised","LOHA: spectral contrastive learning beats supervised on heterophily","Self-supervised LOHA uses low/high-pass views for 77.1% accuracy","Spectral GNN LOHA: 77.1% with no labels, outdoes supervised"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001391,"raw_usage":{"total_tokens":5663,"prompt_tokens":1012,"completion_tokens":4651,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":4573}},"tokens_in":628,"tokens_out":4651,"duration_ms":30685,"temperature":1.0,"reasoning_tokens":4573,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:59:37.097611+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the empirical variance of the composite feature $C_-(z^l_i,z^h_i)$ for fixed nodes across training epochs while the filters update; if it drifts as much as the raw low-pass or high-pass embeddings, the stability premise of Theorem 1 is contradicted and the reunion loss is not acting as the claimed fixed landmark.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the PolyGCL baseline and the low-pass/high-pass filter-contrastive setup that LOHA directly compares to and ablates against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines Dirichlet energy, whose degree-normalized neighborhood difference inspires the spectral signal trend in Eq. 5."},{"cited_title":"V.; and Ye, M","cited_arxiv_id":null,"evidence_quote":"S3GCL is the competing low-pass/high-pass self-supervised framework whose cosine-parameterized filters Eq. 9 extends with a sliding term."}],"review_version":1}