{"id":"27bf3266-a25b-474c-827f-3bfa45d1572c","arxiv_id":"2504.21206","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"FedHERO shares a learned latent-graph generator across federated clients and keeps a private local channel, improving node classification when client graphs have different heterophily patterns.","lead":"This paper proposes FedHERO, a federated learning method for graph node classification that targets heterophilic graphs, where connected nodes often have different labels. It shares a learned structure model across clients while keeping a private local model, and reports accuracy gains over existing federated graph methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Hard top-k latent graph in Eq. (4) is not differentiable as claimed, so the structure learner may receive no gradient and the central sharing mechanism is unsupported as written.","rationale":"The reader's weakest assumption focuses on whether the shared structure learner actually aligns latent neighbor distributions across clients. My primary objection is more fundamental: as specified, the structure learner cannot be trained at all. Eq. (4) selects a hard kNN graph, and the paper explicitly claims differentiability of this operation. Hard top-k is piecewise constant, so gradients do not flow to the parameters that produce the edge scores. Consequently, the mechanism that the paper names as the source of its gains—'By sharing the structure learning model across clients, FedHERO ensures that the latent graphs generated by different clients exhibit similar neighbor distribution patterns'—is unsupported by the stated optimization procedure. This is an internal inconsistency, not a disagreement with external consensus. The paper's own real-world results section (Section 4.3) contains a direct contradiction between 'consistently outperforms all FGL baselines' and 'similar performance between our method and other FGL baselines,' which further weakens the empirical claim. I therefore agree with the reader's REJECT verdict, although I anchor the rejection on the differentiability flaw rather than on the unproven alignment premise. A revised version that discloses and validates a straight-through or soft top-k path, reports tabulated real-world numbers, and tempers the 'consistently outperforms' claim could be reconsidered as CONDITIONAL.","tokens_in":30911,"tokens_out":4364,"duration_ms":51389,"concrete_test":"Attach a gradient hook to the structure learner parameters θ and run one local training step on Squirrel METIS M=7 with the released code. Record ||∂L_total/∂θ|| after the backward pass. If the norm is exactly zero (expected for hard top-k), replace top-k with a straight-through/soft top-k and re-run the main table; if accuracy is unchanged, the reported gains do not depend on learning the structure learner, invalidating the §4.2 explanation. If the norm is nonzero, inspect code to identify the differentiable surrogate and require it to be described in the paper.","verdict_should_be":"REJECT","load_bearing_attack":"Section 3.1 defines the latent graph by hard kNN selection (Eq. 4) and justifies it by 'the differentiability of the top-k function.' This is incorrect: a hard top-k selection is piecewise constant—the entries of \\tilde A are binary indicators of rank position, so ∂\\tilde A_uv / ∂\\tilde a_uv = 0 almost everywhere. Therefore no gradient from L_ce (Eq. 8) reaches the structure learner parameters θ through the latent graph. The regularization L_smooth (Eq. 7) is also computed on the binary \\tilde A, so it too provides zero gradient through the selection. If the implementation instead uses a straight-through estimator or a soft/perturbed top-k, that is an unstated and material deviation from the text. Without a differentiable path, the 'shared structure learning model' is never trained from data; its parameters remain at initialization (modulo FedAvg of identical initializations), and the claimed mechanism—that sharing the learner aligns latent neighbor distributions across clients—cannot be the explanation for the reported gains. This is the load-bearing assumption of the algorithm, not merely a missing proof: the training objective as written does not optimize the structure learner at all. The real-world reporting in Section 4.3 ('consistently outperforms' vs. 'similar performance') is a separate empirical inconsistency, but the gradient issue attacks the core training loop before results are even interpreted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FedHERO, a federated graph learning framework for node classification on heterophilic graphs. Each client maintains a dual-channel GNN: a global channel that operates on a latent graph generated by a structure learner shared across clients, and a private local channel that operates on the original graph. The server aggregates only the global-channel models via FedAvg-style averaging. The paper claims that this design enables clients to share structural knowledge without aggregating conflicting neighbor-distribution patterns, and reports consistent improvements over FGL baselines on semi-synthetic and real-world datasets. The appendices include additional experiments on overlapping subgraphs, latent-graph generation variants, hyperparameter sensitivity, and a detailed comparison with FedStar.","tokens_in":31198,"tokens_out":12976,"duration_ms":119651,"significance":"If the technical concern about differentiability is resolved, the paper makes a useful contribution: a dual-channel architecture with a shared structure learner is a plausible way to reduce cross-client inconsistency in federated learning on heterophilic graphs. The evaluation is broad: four semi-synthetic and three real-world datasets, two partitioning methods, multiple client counts, ablation studies on structure-learning variants and sharing mechanisms, a robustness study, a privacy analysis, and a convergence comparison. The paper also clearly identifies a real problem in FGL. However, because the global channel's claimed training mechanism is not actually optimized as written, these strengths cannot be credited until the method is corrected.","major_comments":[{"comment":"The hard top-k selection in Eq. (4) is not differentiable as claimed. The entries of the binary matrix \\tilde A are piecewise-constant functions of the continuous scores from Eq. (3), so the derivatives of the cross-entropy loss in Eq. (8) and of the smoothness regularizer in Eq. (7) with respect to the structure-learner parameters theta vanish almost everywhere. As written, the structure learner g_L receives no gradient from the training loss, so it cannot be trained from data and the central mechanism of sharing a learned structure learner across clients is unsupported. Please replace Eq. (4) with a differentiable relaxation (for example, a soft top-k with temperature or a perturbed-optimizer estimator) or explicitly describe and analyze a straight-through estimator, and then verify empirically that the structure learner parameters actually change during training.","section":"Section 3.1, Eq. (4), and Appendix C.2"},{"comment":"The text first states that FedHERO 'consistently outperforms all FGL baselines on ieee-fraud and credit datasets' and then immediately attributes 'similar performance' on the same two datasets to the ease of the classification task. These statements are contradictory. The authors should report the real-world results accurately and qualify the 'consistently outperforms' claim accordingly, both in Section 4.3 and in the abstract.","section":"Section 4.3, after Figure 3"},{"comment":"The reported computational cost for FedHERO, O(Ld(d|V|+|E|)+NHd|V|), does not account for constructing the kNN latent graph. Under Eq. (4), every node must be scored against every other node through the metric in Eq. (3), which requires O(NH |V|^2) pairwise computations per client. The complexity table therefore understates the true cost and the claim that FedHERO scales to large graphs 'without introducing much computation burden' is not justified. Please correct the analysis or state an explicit approximation (such as neighbor sampling) that makes the bound valid.","section":"Section 3.4, Table 2"},{"comment":"The assertion that FedHERO 'consistently outperforms all the baseline methods by a significant margin' is stronger than the data in Table 3 supports. For several configurations (for instance, some of the Actor and Flickr columns), the margin over the best baseline is comparable to the reported standard deviation. The empirical claims should be quantified with effect sizes and uncertainty, not just average accuracy.","section":"Section 4.2"}],"minor_comments":[{"comment":"The statement that 'FedAvg often underperforms compared to local training in many experiments' is contradicted by several entries in Table 3 (for example, Actor under METIS M=5, where FedAvg exceeds Local); please soften the claim to reflect the mixed pattern.","section":"Section 4.2"},{"comment":"The same incorrect claim that the top-k function is differentiable appears verbatim in the appendix; any correction in Section 3.1 must be applied here as well.","section":"Appendix C.2"},{"comment":"The link inference attack is cited to Gong and Liu (2018), but that paper describes attribute inference attacks; please provide the appropriate citation for the link inference setting or define the attack precisely.","section":"Section 4.7"},{"comment":"The objective writes g_L(X_i, A_i) without specifying that g_L includes both the one-layer GNN that produces node representations and the metric function of Eq. (3); please make this compositional structure explicit.","section":"Section 2, Eq. (2)"},{"comment":"The notation is hard to follow: the terms f_local, f_global, fl_loc, fl_g, and fc are used interchangeably, and the column headers in Table 3 merge dataset names with M values. Please unify the notation and reformat the table.","section":"Figure 2 and Table 3"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the differentiability claim for Eq. (4). If the authors' code uses a straight-through estimator or a soft top-k, this is a fixable presentation problem; otherwise the paper's central mechanism is not trained at all. I would recommend asking for the code or a clear description of the gradient path before any revision can be evaluated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper takes on a real problem: federated node classification when clients' graphs have different heterophily patterns. The dual-channel idea—share a structure learner whose latent graph feeds a global GNN, keep a local GNN private—is sensible and makes a genuine addition to the FGL toolbox. The semi-synthetic experiments are extensive and the gains over strong baselines (FedStar, FedLit, GCFL) are consistent. The ablations (sharing only the global channel, varying k, structure learner variants) are useful and honest about what matters.\n\nBut there is a load-bearing flaw. Section 3.1 constructs the latent graph by hard kNN selection (Eq. 4) and justifies it by \"the differentiability of the top-k function.\" That is wrong. A hard top-k indicator is piecewise constant; its derivative with respect to the input scores is zero almost everywhere. So neither the cross-entropy loss nor the smoothness regularizer in Eq. (7) provides any gradient to the structure learner parameters. As written, the shared structure learner is never trained from data; it stays at its initial weights, and the claimed mechanism—that aggregating clients' structure learners aligns latent neighbor distributions—cannot be what drives the reported improvements. If the implementation actually uses a soft or straight-through top-k, that is a material deviation from the text and needs to be stated.\n\nThere is also a minor but telling inconsistency in Section 4.3: the text says FedHERO \"consistently outperforms\" on ieee-fraud and credit in one sentence, then attributes its \"similar performance\" to easy features in the next. That needs a fix, but it is secondary.\n\nThe positive side: the problem is well motivated, the experimental design is thorough (METIS and Louvain partitions, three naturally partitioned datasets, multiple baselines, privacy and robustness checks), and the empirical results look credible. The central idea is worth pursuing. But the central mechanism is formally broken as written, and the paper cannot be accepted without a correct differentiable treatment of the latent graph and a clear description of the actual gradient estimator.\n\nWho is this for? Researchers working on federated graph learning or heterophily-aware GNNs will want to know about it. It deserves a serious referee—a skilled reviewer can help the authors fix the gradient issue and tighten the real-world claims. I would not cite it in its current form.\n\nRecommendation: major revision, with the differentiability fix as a precondition.","headline":"Real problem, plausible dual-channel design, but the hard top-k latent graph makes the shared structure learner untrainable as written; the central mechanism needs a differentiable fix.","tokens_in":31754,"tokens_out":3791,"would_cite":false,"duration_ms":38526,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FedHERO claims that sharing a learned latent-graph generator across clients makes federated node classification work on heterophilic graphs, while a private local channel preserves client-specific topology.","keywords":["federated graph learning","heterophilic graphs","node classification","graph structure learning","dual-channel GNN","federated aggregation","privacy-preserving learning"],"falsifier":"Measure the class-mixing pattern of the latent graphs each client produces after federation, for example by computing the Jensen-Shannon divergence between neighbor-class distributions of latent graphs across client pairs; if those divergences are not smaller than the same divergences computed on the original local graphs, the shared structure learner is not aligning the clients and the reported gains must come from some other component. Alternatively, train FedHERO with the shared structure learner but randomize the server-side averaging of global channels and check whether the accuracy gap to FedAvg persists.","tokens_in":30668,"feed_emoji":"🕸️","tokens_out":8528,"duration_ms":78758,"temperature":0.7,"pith_summary":"Federated graph learning usually assumes that all clients' graphs are homophilic, so that the GNNs trained on them capture compatible patterns and can be safely averaged. This paper argues that the assumption fails in practice, because clients may hold graphs with different heterophily levels, and standard aggregation such as FedAvg makes the global model worse than local training. To fix it, the paper proposes FedHERO, which splits each client model into a shared global channel driven by a learned latent graph and a private local channel that operates on the original graph, and claims this design consistently outperforms existing federated graph learning baselines on node classification. If that claim holds, federated node classification no longer needs to assume homophily to benefit from collaboration.","feed_headline":"Federated graph learning beats heterophily by sharing latent graphs","feed_subtitle":"A shared structure learner aligns client graphs while private local channels keep topology local, and the method tops six federated…","key_machinery":"The load-bearing object is the dual-channel GNN with a shared structure learner. The structure learner is a one-layer GNN plus a multi-head weighted-attention metric $\\phi(z_u,z_v) = \\frac{1}{N_H}\\sum_h \\cos(w_h^1 \\odot z_u, w_h^2 \\odot z_v)$ that scores each candidate edge, and a top-$k$ selection turns the scores into a latent graph $\\tilde{A}$. The global channel propagates on $\\tilde{A}$, the local channel propagates on the original $A$, and the layer output is a convex combination $Z^\\ell = \\sigma(\\alpha E^\\ell + (1-\\alpha) H^\\ell)$. Only the global-channel parameters $w_g$ are aggregated as $w_g = \\sum_i \\frac{N_i}{N} w_i^g$, which is what lets clients share a common message-passing pattern while keeping their own topology private.","core_discovery":"The central claim is that the obstacle to federated learning on heterophilic graphs is not heterophily itself but the mismatch between neighbor-distribution patterns across clients, and that this mismatch can be removed by learning structure. FedHERO gives each client a structure learner that scores edges between nodes with a multi-head weighted attention metric, applies a top-k sparsification to build a latent graph, and then runs a global-channel GNN on that latent graph while a local-channel GNN runs on the original adjacency. Only the global channel parameters are sent to the server and averaged by node-count weighting; the local channel and classifier stay private. The authors assert that the shared structure learner makes the latent graphs from different clients exhibit similar neighbor-distribution patterns, so the aggregated global-channel GNNs capture consistent knowledge, and that FedHERO consistently outperforms all tested federated graph learning baselines on four semi-synthetic heterophilic datasets and three real-world datasets.","pith_inferences":["The same dual-channel recipe could transplant to other federated settings where clients share a task but disagree about the local geometry of the input space: learn a shared canonical neighborhood structure, aggregate only the models trained on it, and keep the local geometry private.","The paper's explanation implies a testable monotonicity: across a synthetic sweep of cross-client neighbor-distribution mismatch, FedHERO's advantage over FedAvg should widen as the mismatch grows; such a sweep would isolate the mechanism from the architecture's other benefits.","The privacy claim is bounded by the attack tested: sharing a structure learner that maps features to edge scores may still leak topology through the shared parameters, so the reported privacy gain is relative to link inference from node representations rather than a formal guarantee.","The top-k post-processing may be doing more than sparsifying: the Bernoulli-sampling variant performs worse and less stably, suggesting that the fixed, discrete degree of the latent graph, not just the learned edge scores, contributes to cross-client alignment."],"forward_implications":["Federated node classification no longer has to assume client graphs are homophilic; clients with different heterophily levels can still get a benefit from aggregation.","Only the global channel is shared, so client-specific topology and the classifier stay private; the paper's link-inference attack results show that sharing this smaller part leaks less structural information than sharing whole models.","Because message passing runs on a learned latent graph rather than the raw adjacency, flipping ten percent of edges degrades FedHERO less than it degrades the baselines.","The structure learner is a swappable module: attention, GAT, cosine distance, GLCN, and GAug variants all work, with the multi-head attention version performing best in the paper's comparisons.","The performance gap over the best baseline grows on datasets with lower homophily ratios, indicating that the method targets exactly the regime where existing federated graph learning methods fail."],"supporting_citations":[{"why":"Defines the FedAvg aggregation rule that FedHERO adapts, averaging only global-channel parameters.","marker":"McMahan et al., 2017"},{"why":"Supplies the multi-head weighted attention metric used as the structure learner's edge scorer.","marker":"Chen et al., 2020"},{"why":"Provides heterophilic GNN design principles and the syn-cora dataset used in evaluation.","marker":"Zhu et al., 2020"},{"why":"FedStar is the closest baseline and the structure-sharing method FedHERO contrasts with.","marker":"Tan et al., 2023"},{"why":"FedPub is a key baseline, and its subgraph partition methodology is reused for the federated datasets.","marker":"Baek et al., 2023"},{"why":"FedSage is a baseline and supplies the missing-neighbor-generation idea in node-level federated graph learning.","marker":"Zhang et al., 2021"},{"why":"FedLit is a baseline that models latent link types, the closest alternative in message-passing adaptation.","marker":"Xie et al., 2023"},{"why":"Provides the Squirrel and Chameleon heterophilic datasets used in the main evaluation.","marker":"Rozemberczki et al., 2021"}],"fun_headline_variants":["Shared latent graphs let federated GNNs master heterophily","FedHERO: federated GNNs learn structure to handle heterophily","Aligning neighbor patterns with a shared structure learner","Federated GNNs share latent graphs to overcome heterophily"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the premise that averaging locally trained structure learners makes the latent graphs they generate similar enough across clients that aggregating the global-channel models helps; if the latent graphs stay as mismatched as the original graphs, the whole benefit disappears.","fun_headline_variants_meta":{"raw":{"variants":["Shared latent graphs let federated GNNs master heterophily","FedHERO: federated GNNs learn structure to handle heterophily","Aligning neighbor patterns with a shared structure learner","Federated GNNs share latent graphs to overcome heterophily"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000486,"raw_usage":{"total_tokens":2433,"prompt_tokens":1017,"completion_tokens":1416,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":633,"completion_tokens_details":{"reasoning_tokens":1341}},"tokens_in":633,"tokens_out":1416,"duration_ms":9754,"temperature":1.0,"reasoning_tokens":1341,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:10:26.888433+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the class-mixing pattern of the latent graphs each client produces after federation, for example by computing the Jensen-Shannon divergence between neighbor-class distributions of latent graphs across client pairs; if those divergences are not smaller than the same divergences computed on the original local graphs, the shared structure learner is not aligning the clients and the reported gains must come from some other component. Alternatively, train FedHERO with the shared structure learner but randomize the server-side averaging of global channels and check whether the accuracy gap to FedAvg persists.","supporting_citations":[{"cited_title":"Federated node classification over graphs with latent link-type heterogeneity","cited_arxiv_id":null,"evidence_quote":"FedLit is a baseline that models latent link types, the closest alternative in message-passing adaptation."}],"review_version":1}