{"id":"a00ed6c9-6259-413a-ab67-8979fd5e98ea","arxiv_id":"2411.11304","paper_version":7,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"O-pFGL achieves one-shot personalized federated graph learning by aggregating class-wise feature statistics into a surrogate graph and combining global distillation with local fine-tuning, outperforming baselines on 14 graph datasets.","lead":"This paper proposes O-pFGL, a method for personalized federated graph learning that requires only one round of communication. Clients share class-wise feature statistics instead of model parameters, the server builds a small 'surrogate graph', and each client fine-tunes a global model on its own data with distillation to improve minority-class accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 8's global variance estimator is not unbiased: it divides total sum of squares by N_c - m, so even under IID it overestimates by (N_c-1)/(N_c-m), and for minority classes with N_c < m it can become negative; this corrupts the surrogate graph that the whole method relies on.","rationale":"I considered the reader's stated weakest assumption, that per-class first and second moments of propagated features are sufficient to represent the global graph. That is a genuine limitation and is worth flagging, but it is not the most sharply checkable defect in the paper. The concrete, demonstrable problem is Eq. 8: the formula aggregates class-wise variances in a way that is mathematically not unbiased under any standard interpretation. The numerator is the total sum of squares, whose unbiased estimator divides by N_c - 1, while the denominator N_c - m corresponds to the pooled within-group degrees of freedom. The paper's own text claims the recovered statistics remain unbiased regardless of non-IIDness, and this is directly falsified by taking expectations. Because the global surrogate graph is built by matching these statistics (Eq. 10), an overdispersed or negative variance target degrades the fidelity of the surrogate, especially for small classes—the very classes central to the paper's F1-macro improvements. The empirical evaluation is broad and appears clean on label leakage; the ablation and scalability studies are supportive, so I do not see grounds to reject the paper. However, the variance aggregation error is a real correctness defect that should be fixed and the key tables re-run before full acceptance. This matches the reader's CONDITIONAL verdict, so no change is needed.","tokens_in":48186,"tokens_out":9361,"duration_ms":94347,"concrete_test":"Analytically compute E[s²_c] from Eq. 8 under a one-way random-effects model (client means μ_k ~ N(μ, τ²), features x|k ~ N(μ_k, σ²)) and compare with the unbiased total variance. Then re-run the full O-pFGL pipeline on Cora and ogbn-arxiv with the corrected estimator s²_c = S_total/(N_c - 1) (or, if within-client variance was intended, S_within/(N_c - m)) and report accuracy/F1-macro. If the corrected results differ by more than a small margin, or if any class yields N_c < m, Eq. 8 is load-bearing; if results are essentially unchanged, the flaw is a theoretical defect rather than the driver of the empirical claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The weakest load-bearing point is the aggregation formula for the global class variance, Eq. 8. The numerator is the total sum of squares S_total = Σ_k (N_k^c - 1)s_k² + Σ_k N_k^c (μ_k^c - μ_c)². The unbiased estimator of the total (mixture) variance is S_total/(N_c - 1); the unbiased estimator of within-client variance is the first term divided by Σ_k(N_k^c - 1), i.e., N_c - m only when every client contributes. Eq. 8 instead divides S_total by N_c - m. Even when all clients are IID draws from a single distribution, E[S_total] = (N_c - 1)σ², so E[s²_c] = (N_c - 1)/(N_c - m) σ², which is an upward bias; with heterogeneous client means the bias is larger. For a minority class with total count N_c < m, which cannot be ruled out under the label-skew partitions used here, the denominator is zero or negative and the formula yields an undefined or negative variance unless an unstated fallback exists. Since this variance is exactly the target that the global surrogate graph optimizes in Eq. 10, the recovered 'global distribution' is distorted precisely for the minority classes the method claims to protect. The statement that the recovered statistics 'remain unbiased regardless of the degree of data non-IIDness' is therefore false. This error does not by itself prove the reported accuracy/F1 gains are spurious, but it removes the core theoretical guarantee and can materially change the surrogate graph.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes O-pFGL, a one-shot personalized federated graph learning method for node classification. Each client estimates class-wise means and variances of propagated node features and uploads these statistics in a form compatible with Secure Aggregation. The server aggregates the statistics into a purported global class-wise distribution and synthesizes a small global surrogate graph by matching the first two moments of propagated features and enforcing a smoothness prior. Clients then train a generalized model on the surrogate graph (Stage 1) and fine-tune it on local data with node-adaptive knowledge distillation that up-weights minority and low-homophily nodes (Stage 2). The method supports heterogeneous client models. Experiments on 14 real-world datasets compare O-pFGL against 11 baselines under Louvain/Metis partitions with 10 and 20 clients, plus heterophilic, inductive, model-heterogeneity, and large-client settings. The paper reports consistent accuracy and F1-macro gains, and Appendix D.2 reports better average performance than multi-round baselines run for 100 rounds.","tokens_in":48414,"tokens_out":8103,"duration_ms":79915,"significance":"The problem is timely and the empirical study is unusually broad: 14 datasets, multiple partition schemes, 10/20 clients, inductive and heterophilic benchmarks, model-heterogeneity experiments, and a detailed ablation study. The evaluation is on held-out nodes and the surrogate graph is generated from training-set statistics, so the reported gains are not fitted values; the empirical contribution is genuine evidence for the approach. The communication cost (class-wise statistics plus a small surrogate graph) and the compatibility with Secure Aggregation are practical strengths. However, the paper's central theoretical claim that the aggregated global variance is unbiased is mathematically incorrect, and the implementation of the variance alignment for surrogate graphs with one node per class is left unspecified. The main result is therefore supported empirically but not by the stated derivation; the manuscript needs a corrected aggregation formula and a corresponding re-examination of the surrogate-generation loss before the claims can be accepted.","major_comments":[{"comment":"The aggregation formula in Eq. (8) is not an unbiased estimator of the global class-wise feature variance, contradicting the sentence immediately after it. Under IID sampling from a common distribution per class, the total sum of squares in the numerator has expectation (N_c - 1) sigma_c^2, so dividing by N_c - m yields E[s^{2c}] = (N_c - 1)/(N_c - m) sigma_c^2, which overestimates the variance. Under heterogeneous client means, the between-client term additionally inflates the estimate, so the bias grows with non-IIDness. Moreover, if a class has N_c < m total samples, which can occur under the label-skew partitions used in the experiments, the denominator is non-positive and the formula yields an undefined or negative variance with no stated fallback. Since Eq. (10) aligns the surrogate graph's variance to this value, the distortion propagates into the global model and the distillation stage. The manuscript should either divide by N_c - 1 to estimate the pooled total variance or explicitly define and justify the target quantity being estimated, and it must specify a valid fallback for small N_c.","section":"Global Surrogate Graph Generation, Eq. (8)"},{"comment":"The surrogate graph generation uses a sample variance defined in Eq. (5) with denominator |V| - 1. Appendix C states that on Cora, CiteSeer, and PubMed the global surrogate graph is set to exactly one node per class. For a class with one node, the sample variance is undefined (division by zero), yet Eq. (10) includes the term (s'^{2c} - s^{2c})^2. The paper does not describe the fallback estimator actually used in the implementation, nor how the variance alignment is computed for one-node classes. This is a load-bearing implementation detail because the entire surrogate graph is generated from these moments. The text must state the actual variance estimator used for the surrogate graph or adjust the node-count configuration so that the stated equations are well defined.","section":"Global Surrogate Graph Generation, Eq. (5) and Appendix C"},{"comment":"The surrogate graph is generated by matching only the first two moments of propagated features plus a smoothness prior. The paper does not provide evidence that these statistics capture the structural information needed to transfer global knowledge in heterophilic or structurally heterogeneous graphs, such as degree distributions, edge homophily, or higher-order neighborhood structure. Because the surrogate graph is the only channel through which cross-client knowledge enters Stage 1 and the distillation stage, a diagnostic test would substantiate the mechanism: for example, compare the surrogate graph's edge homophily and degree distribution with the client graphs, or compare the Stage-1 model's accuracy against a model trained on an oracle global graph. Without such evidence, the end-to-end gains demonstrate that the method works but do not establish that the proposed surrogate-graph construction is the reason.","section":"Global Surrogate Graph Generation, Eqs. (10)-(11)"}],"minor_comments":[{"comment":"The local class-wise variance is undefined when a client has exactly one labeled node of a class, since the denominator |V^c_{k,L}| - 1 is zero. The manuscript says the statistics are estimated only for classes with sufficient labeled nodes but never defines the minimum count or a fallback estimator.","section":"Feature Distribution Estimation, Eq. (5)"},{"comment":"The Actor row lists the property as 'Heterophlic'; this should be 'Heterophilic'.","section":"Table 3"},{"comment":"The sentence 'We set the communication round of these 6 methods to 100' is followed by a list of seven methods (FedAvg, FedPUB, FedGTA, FedTAD, FedSpray, FedLoG, and FedGM); the count should be corrected.","section":"Appendix D.2"},{"comment":"The definition of hnode(vi) divides by |N_vi ∩ V_L|, which is zero for an isolated labeled node; the paper should state the convention adopted in that case.","section":"Preliminaries, Eq. (1)"},{"comment":"The statement that the aggregation process is 'lossless and unbiased regardless of the data distributions' repeats the incorrect claim from Eq. (8) and should be revised consistently with the corrected variance estimator.","section":"Appendix J"},{"comment":"The text describes the pipeline as comprising five steps, but the figure does not show step indices; adding visible step numbers would make the correspondence between the text and the figure much clearer.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The variance-estimator error in Eq. (8) is concrete and fixable, but the paper's repeated 'unbiased' claims (Eq. (8) and Appendix J) must be corrected before acceptance, and the one-node-per-class surrogate configuration in Appendix C needs a described variance fallback. The empirical comparison is extensive and mostly well executed; the two-stage distillation and the low-communication design are valuable contributions. One additional concern for the editor: the paper does not mention code or data release, which would strengthen reproducibility given the large number of reported tables and tuning details. The comparison with 100-round baselines in Appendix D.2 is a strong point, but the average rows are computed over slightly different dataset subsets because some baselines run out of memory, so the average comparison should be framed with that caveat."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the first paper I know that combines one-shot communication with personalization for federated node classification. The empirical sweep is unusually wide—14 datasets, Louvain/Metis partitions, 10/20/1000 clients, homophilic/heterophilic, inductive, heterogeneous models, several GNN backbones—and the gains over the next-best baselines, especially in F1-macro, are consistent. The two-stage node-adaptive distillation is a sensible mechanism: train a global teacher on a server-side surrogate graph, then fine-tune locally with per-node KL weighting based on class homophily. HRE is a reasonable plug-in. The evaluation is on held-out nodes, so the result is not a circular fit. That part holds up.\n\nThe soft spot is real and load-bearing. Eq. 8 divides the total sum of squares by N_c - m and calls the result an unbiased global variance. That is not unbiased. The unbiased estimator of the total variance is S_total/(N_c - 1); dividing by N_c - m biases it upward by (N_c-1)/(N_c-m), and when a minority class has fewer total samples than there are clients, the denominator is non-positive. This is exactly the target the surrogate graph aligns to in Eq. 10, so the recovered global distribution is distorted for the minority classes the paper most wants to protect. The claim in Eq. 8 and repeated in Appendix J that the recovered statistics are unbiased regardless of non-IIDness is simply false.\n\nDoes that kill the empirical result? Not automatically—the reported gains could still hold because the optimization relaxes the alignment loss, and the method may work fine with a biased variance target. But it removes the paper's main theoretical guarantee, and the method is otherwise built on an unproven sufficiency assumption: that class-wise means and variances of propagated features capture enough of the global graph structure. That assumption is not justified.\n\nAlso: no code is released, and the per-dataset hyperparameters (p, f_th, d_th, K, beta, delta, alpha) are many and tuned per dataset. The appendix gives sensitivity for some, but an off-the-shelf deployment story is weak.\n\nWho this is for: anyone working on one-shot federated learning or personalized FGL. The problem framing is valuable and the empirical evidence is worth taking seriously. A serious referee should get this paper, but the authors need to correct the variance claim, release code, and either prove or drop the sufficiency assumption. I'd send it out rather than desk-reject.","headline":"First real one-shot personalized FGL paper with broad, consistent empirical gains, but the headline unbiasedness claim for its variance aggregation is wrong and needs fixing before the method's guarantees can be taken at face value.","tokens_in":49116,"tokens_out":2607,"would_cite":true,"duration_ms":28370,"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":"One-shot federated graph learning beats 100-round methods on average, the paper claims.","keywords":["one-shot federated learning","federated graph learning","personalized federated learning","node classification","surrogate graph","knowledge distillation","label propagation","data heterogeneity"],"falsifier":"Run O-pFGL on a synthetic heterophilic graph whose class-wise feature means and variances are constructed to be identical across classes while labels are determined by triadic closure or other higher-order patterns. If the method fails to beat standalone local training on that graph, the claim that class-wise first and second moments of propagated features suffice for a useful surrogate graph is refuted.","tokens_in":47797,"feed_emoji":"🕸️","tokens_out":5180,"duration_ms":47367,"temperature":0.7,"pith_summary":"This paper tries to establish that personalized federated learning on graphs can be done in a single communication round without sacrificing accuracy or fairness. The proposed method, O-pFGL, has each client estimate class-wise means and variances of propagated node features, the server combines these statistics into a small global surrogate graph, and each client then trains its own model in two stages with a distillation step that protects minority classes. The claim is that this one-shot procedure outperforms existing federated graph learning and one-shot federated learning baselines across 14 datasets, and even beats multi-round methods given 100 rounds. If true, it would make collaborative graph learning practical where communication is expensive or risky, and would let clients keep both their data and their model architectures private.","feed_headline":"One-shot graph federated learning beats 100-round rivals","feed_subtitle":"Clients share only class-wise statistics; a server-built surrogate graph yields balanced personalized models in one round.","key_machinery":"The load-bearing object is the global surrogate graph $G' = (A', X', Y')$, a small graph (under 300 nodes in the experiments) generated on the server so that its class-wise propagated feature means and variances match the unbiased aggregated global statistics. The adjacency matrix $A'$ comes from a learnable link predictor $g_\\theta$ over node features, with a smoothness loss $\\mathcal{L}_{\\text{smt}}$ to encourage a realistic structure, and the node features $X'$ are optimized against the alignment loss $\\mathcal{L}_{\\text{align}}$. Around this object sit two supporting mechanisms: the Homophily-guided Reliable node Expansion (HRE) strategy, which uses label propagation and class homophily to expand scarce labeled sets for more accurate statistics, and node-adaptive distillation, which weights the KL divergence from the global teacher by $\\gamma_i = \\beta \\tilde{y}_i \\cdot w_{\\text{dist}}$ with $w_{\\text{dist}}[c] = 1/(1 + \\log(H(c) + 1))$, so that nodes in low-homophily or minority classes inherit more global knowledge.","core_discovery":"The central discovery is that the information needed to personalize a graph model across heterogeneous clients can be carried by class-wise first and second moments of propagated features, rather than by model parameters or raw graph data. From these aggregated moments the server synthesizes a small surrogate graph whose propagated feature distribution matches the recovered global distribution, and this surrogate graph serves as a teacher for a two-stage local training procedure. In the second stage, a node-adaptive distillation weight, derived from class homophily and soft labels, decides per node how much to trust the global teacher versus local fine-tuning, which the paper argues corrects the bias of fine-tuning toward majority classes. The paper reports consistent gains in accuracy and F1-macro over eleven baselines on 14 real-world datasets, and better average performance than multi-round methods allowed 100 communication rounds.","pith_inferences":["The method's success would imply that cross-client graph knowledge transfer can be reduced to moment matching of propagated features, a much weaker commitment than sharing structure; this suggests a general design principle for one-shot federated learning on non-Euclidean data.","A testable extension is to apply the surrogate-graph idea to graph-level tasks (graph classification) or to link prediction, where the same statistics-plus-smoothness recipe might need higher-order moments to capture motif structure.","The reliance on class-wise moments predicts a failure mode: on graphs where classes are separated by higher-order structural patterns (e.g., role detection in heterophilic networks), the surrogate graph may be information-poor; experiments that withhold second-order moments or scramble topology would isolate how much structure the method actually uses.","Because the method assumes honest clients and an honest server, an adversary-free setting is load-bearing for the privacy claims; a malicious server could still infer class proportions and feature statistics from the aggregated moments."],"forward_implications":["Communication cost becomes independent of model size: clients upload only $O(C \\cdot h \\cdot f)$ statistics and download a small surrogate graph, which scales better than parameter exchange as models grow.","Because no model parameters are shared, clients can use heterogeneous architectures and protect model intellectual property, with statistics upload compatible with Secure Aggregation protocols.","The node-adaptive distillation is claimed to raise F1-macro substantially over fine-tuning, implying better performance on minority classes in imbalanced local graphs.","The method reportedly beats multi-round federated baselines trained for 100 rounds on average, suggesting one-shot collaboration may be sufficient for node classification in many real-world regimes.","The approach carries over to inductive datasets and other GNN backbones (GraphSage, GAT, SGC) per the experiments."],"supporting_citations":[{"why":"Supplies the MixHop-style concatenation of multi-hop propagated features that makes the class-wise estimation work on heterophilic graphs.","marker":"(Abu-El-Haija et al. 2019)"},{"why":"Provides the alignment-loss recipe used to optimize the surrogate graph's features against target class-wise statistics.","marker":"(Xiao et al. 2024)"},{"why":"Provides label propagation, which generates the soft labels used both in HRE expansion and in computing node-adaptive distillation weights.","marker":"(Iscen et al. 2019)"},{"why":"GHOST is the closest one-shot FGL baseline, whose limitations motivate O-pFGL and which the paper's experiments show it outperforms.","marker":"(Qian et al. 2025)"},{"why":"The secure aggregation protocol whose weighted-average structure the upload-and-aggregate statistics are shown to be compatible with.","marker":"(Bonawitz et al. 2017)"},{"why":"FedPUB is a representative multi-round personalized FGL method used as a baseline and as evidence that personalization has required many rounds.","marker":"(Baek et al. 2023)"},{"why":"DENSE is a one-shot FL baseline adapted for graphs, illustrating why image-oriented one-shot methods fail on interdependent graph data.","marker":"(Zhang et al. 2022)"}],"fun_headline_variants":["One-shot graph federated learning: class stats build teacher graph","One-round personalized graph FL via surrogate graphs","Graph FL in one shot: client stats create global teacher","One-shot personalized FGL: surrogate graph from class moments"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method works only if a small graph whose class-wise propagated feature means and variances match the global statistics captures enough of the true graph structure and feature distribution to train a teacher model that improves each client's personalization.","fun_headline_variants_meta":{"raw":{"variants":["One-shot graph federated learning: class stats build teacher graph","One-round personalized graph FL via surrogate graphs","Graph FL in one shot: client stats create global teacher","One-shot personalized FGL: surrogate graph from class moments"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000382,"raw_usage":{"total_tokens":2041,"prompt_tokens":980,"completion_tokens":1061,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":996}},"tokens_in":596,"tokens_out":1061,"duration_ms":8151,"temperature":1.0,"reasoning_tokens":996,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:40:56.608060+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run O-pFGL on a synthetic heterophilic graph whose class-wise feature means and variances are constructed to be identical across classes while labels are determined by triadic closure or other higher-order patterns. If the method fails to beat standalone local training on that graph, the claim that class-wise first and second moments of propagated features suffice for a useful surrogate graph is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the MixHop-style concatenation of multi-hop propagated features that makes the class-wise estimation work on heterophilic graphs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the alignment-loss recipe used to optimize the surrogate graph's features against target class-wise statistics."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides label propagation, which generates the soft labels used both in HRE expansion and in computing node-adaptive distillation weights."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GHOST is the closest one-shot FGL baseline, whose limitations motivate O-pFGL and which the paper's experiments show it outperforms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DENSE is a one-shot FL baseline adapted for graphs, illustrating why image-oriented one-shot methods fail on interdependent graph data."}],"review_version":1}