{"id":"fe496b11-c994-4720-a238-9896cfd01a94","arxiv_id":"2412.13442","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"CEFGL combines low-rank global and sparse personalized model components with quantization and communication skipping to improve accuracy and cut communication cost in federated graph classification.","lead":"A federated graph learning method that splits each client model into a shared low-rank part and a private sparse part, then compresses communication with quantization and occasional server skips. It reports higher accuracy than prior federated graph methods on cross-dataset benchmarks while sending far fewer bits.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's update equations are dimensionally invalid and the skip-communication branch is incoherent, so the reported accuracy and communication gains cannot be reproduced from the paper as written.","rationale":"The paper's advertised contribution is an algorithm with a specific optimization procedure and specific measured gains. For the central claim to hold, the implemented method must be the one described in Section 4 and Algorithm 1. The description is internally broken in ways that prevent a reader from even executing the update rules: Eq. (6) and Eq. (8) add scalar norms to parameter matrices, which is not a valid operation in the asserted update, and the skip-communication branch cannot be implemented as written because it depends on a client-local variable that the server never receives. These are not stylistic issues; they directly undermine reproducibility of the experimental results. The ablation table adds a second, independent inconsistency: the full CEFGL model in Table 5 should coincide with the 'Ours' row in Table 1 for the same backbone and setting, but the CHEM values differ by 2.4 accuracy points (83.03 vs 85.43). This suggests the numbers in the main tables come from a variant that is not fully specified. No code is provided, so the discrepancy cannot be resolved by inspecting the artifacts. I am not claiming the results are fabricated; the method may work well in practice. But the paper, as written, does not provide enough information to verify the central empirical claim, and the internal contradictions make it more likely that the described algorithm is not the one that produced the headline numbers. A conditional verdict is therefore appropriate: the idea is plausible, but acceptance should require a corrected, executable description and a resolution of the Table 1/Table 5 mismatch. The reader's weakest assumption correctly identifies the faithfulness of the update rules as the key vulnerability; I agree with that assessment and add the ablation inconsistency as concrete evidence that the assumption fails.","tokens_in":18732,"tokens_out":5175,"duration_ms":47316,"concrete_test":"Obtain the authors' code (or reimplement CEFGL strictly from Eqs. (6)-(9) and Algorithm 1) and run the CHEM cross-dataset GIN configuration with the paper's defaults (mu=0.0001, lambda=0.001, alpha=0.6, p=0.5, 200 rounds). First, verify that Eq. (6) and Eq. (8) are executable as written; then compare the resulting CHEM accuracy with Table 1 (85.43+/-1.14) and Table 5 (83.03). If the literal equations cannot be executed, or the run does not match either number, the paper's reported results do not follow from the described method and the central claim is unverified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: CEFGL outperforms FedStar in accuracy and communication cost. The most load-bearing condition is that the algorithm actually executed matches the equations and pseudocode. That condition fails as written. Eq. (6) is W_i^{t+1}=W_i^t - eta*(grad_W f_i(W_i^t;D_i)-h_i^t) + (alpha/2)*||Theta^t - W_i^t||; the last term is a scalar norm added to a parameter matrix, so the update is dimensionally invalid (the intended proximal term must be a matrix, e.g. (alpha/2)*(Theta^t-W_i^t) or a squared norm inside the loss). Eq. (8) has the same defect: '+||S_i^t||_1' adds a scalar to a matrix. Algorithm 1 line 14 sets Theta^{t+1}=W_i^{t+1} when j=0, but W_i^{t+1} is a client-local variable that is not transmitted on a skipped round; if the branch means 'no communication', the server cannot set Theta to W_i. The inconsistency is not merely cosmetic: Table 5's full-model row (W and S, GIN) reports CHEM=83.03, while Table 1's GIN 'Ours' reports CHEM=85.43+/-1.14 for the same setting and backbone. The ablation's full model should match the main result; it does not. Therefore the 5.64% accuracy gain and 18.58x communication reduction cited in the abstract are not tied to a well-defined, reproducible algorithm.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes CEFGL, a personalized federated graph learning algorithm that decomposes client model parameters into a low-rank shared component and a sparse private component, combines local training with correction terms, uses probabilistic communication skipping and quantization to reduce communication cost, and evaluates the method on sixteen graph-classification datasets with GCN, GIN, and GraphSage backbones. The central empirical claim is that CEFGL outperforms FedStar and other federated graph learning baselines in accuracy while substantially reducing communication bits and time; the headline example is a 5.64 percentage point accuracy improvement on the CHEM cross-dataset setting with GIN, together with an 18.58x reduction in communication bits and a 1.65x reduction in communication time.","tokens_in":19164,"tokens_out":6429,"duration_ms":56219,"significance":"If the claims hold, the paper would make a useful empirical contribution: a single framework that handles non-IID graph data by combining sparse personalization with low-rank global sharing, while reducing communication through quantization and less frequent synchronization. The experimental comparison is extensive, covering three GNN architectures, multiple FGL baselines, sensitivity analyses, dropout robustness, and an ablation of the two model components; these are genuine strengths. However, the load-bearing algorithmic description is not internally consistent as written: the update equations contain dimensionally invalid terms, the skip-communication branch of the pseudocode is incoherent, and the main results are not matched by the ablation table. The empirical claims are therefore not yet tied to a well-defined, reproducible algorithm.","major_comments":[{"comment":"Equation (6) is dimensionally invalid: the term (α/2)||Θ^t − W_i^t|| is a scalar norm added to the parameter matrix W_i^t, so the right-hand side is not a matrix update of the same shape as W_i^t. Equation (8) has the same defect, since +||S_i^t||_1 adds a scalar to the matrix S_i^t. If the intended updates are proximal-style corrections, the equations must be rewritten (for example, with (α/2)(Θ^t − W_i^t) or with the norm placed inside the objective and its gradient taken), and Algorithm 1 lines 6 and 8 must be made consistent. As written, the update rules in Algorithm 1 cannot be executed, which blocks reproduction of the reported accuracy and communication results.","section":"§4.1, Eq. (6) and Eq. (8)"},{"comment":"The skip-communication branch sets Θ^{t+1} = W_i^{t+1} when j = 0, but W_i^{t+1} is a client-local variable that is not transmitted to the server on a skipped round. The branch is also outside the client loop, so the index i is undefined at that point. Choosing a single client's model as the global model would contradict the idea of skipping communication and would likely destroy shared knowledge. The pseudocode needs to specify what the server and clients actually do on skipped rounds (for example, Θ^{t+1} = Θ^t while clients continue local optimization), and the client-selection logic in lines 3–4 needs to be reconciled with the communication-skip decision.","section":"Algorithm 1, line 14"},{"comment":"The full-model ablation row (W and S, GIN) reports CHEM = 83.03, whereas Table 1 reports GIN 'Ours' CHEM = 85.43 ± 1.14 for the same setting and backbone, and Table 4 reports CHEM = 85.03 for the stated default sparsity regime. Since the ablation's full configuration should coincide with the main result, this discrepancy suggests either different hyperparameters, different communication probability, a different number of rounds, or an error. The authors need to explain this mismatch before the headline 5.64% improvement over FedStar can be evaluated.","section":"Table 5 vs. Table 1 and Table 4"},{"comment":"The role of W_i^t and S_i^t in the forward model is ambiguous. Equation (6) updates W_i using a loss f_i(W_i^t; D_i), while Eq. (8) updates S_i using a loss f_i(Θ^t + S_i^t; D_i), and the Figure 1 caption says that during local training the model parameters W_i^t are frozen while the sparsification components are fine-tuned. Remark 1 further states that W_i is not actually low-rank during local training. The paper needs to state the exact forward model and the exact order of operations (simultaneous or sequential training of W_i and S_i, and which parameters are frozen at each stage) so that Algorithm 1 matches the implementation used in the experiments.","section":"§4.1, Figure 1, and Remark 1"}],"minor_comments":[{"comment":"The title uses 'Federal Graph Learning' while the body uses 'Federated Graph Learning'; the title should be corrected to 'Federated'.","section":"Title and Abstract"},{"comment":"The text mentions a baseline 'FedPerGCN' that is not listed in the baseline enumeration in §5.1; this is likely a typo for FedGCN or FedPer and should be corrected.","section":"§5.2"},{"comment":"The quantizer Q_r is described as involving independent random variables, but Eq. (11) gives a deterministic rounding formula. Please either provide the actual probability distribution (as in QSGD) or state explicitly that deterministic rounding is used, since the current definition is ambiguous.","section":"§4.3, Eqs. (10)–(11)"},{"comment":"The text lists the IID single-dataset setups as DD, NCI1, and IMDB-MULTI, but Table 2's third column is IMDB-BINARY; please reconcile the dataset name.","section":"§5.1 and Table 2"},{"comment":"Equation (9) uses Θ^{t+1}_i on the left-hand side for what should be a single global model, and it does not show the quantization applied in Algorithm 1 line 12; please correct the notation and specify whether quantization is applied after the low-rank projection.","section":"§4.2, Eq. (9)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a clearly identified empirical contribution and the experimental breadth is commendable, but the algorithmic description contains load-bearing inconsistencies that make the experiments unreproducible as written. I recommend major revision rather than rejection because the problems are in principle fixable by rewriting the equations and pseudocode, clarifying the skipped-communication protocol, and reconciling the ablation table with the main results. Given that the central claim is empirical, I would also encourage the authors to release code or a detailed appendix with the corrected update rules."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Liu et al. propose CEFGL, a personalized federated graph learning method that decomposes each client model into a low-rank global component and a sparse personal component, then applies quantization and probabilistic communication skipping. The specific integration is new to FGL; the ingredients are borrowed from RPCA, FedRep/pFedMe, QSGD, and Scaffold/ProxSkip, but the recipe is not in the prior work. The experimental footprint is genuinely large: sixteen datasets, three GNN backbones, seven baselines, plus ablations, sensitivity analyses, and communication-cost plots. In the main tables CEFGL consistently tops the average accuracy across settings, matching the communication reductions claimed in the abstract.\n\nThe soft spots are load-bearing. Eq. (6) and Eq. (8) as printed add scalar norms to matrix updates, which is dimensionally invalid; the intended proximal terms are not stated. Algorithm 1 line 14 sets Theta^{t+1} = W_i^{t+1} on a skipped round, but W_i is local and never uploaded, so the server cannot do that. And Table 5's full-model row gives CHEM=83.03 while Table 1's GIN row gives 85.43 for what should be the same configuration. These are not cosmetic typos: they sit exactly where the reader needs to know what the algorithm actually did to trust the headline 5.64% accuracy gain and 18.58x communication reduction. There is also a stray internal name, 'SDGRL', in Section 5.9 that looks like a leftover from an earlier draft.\n\nI would not defend the paper as it stands. The empirical results are plausible but not reproducible from the description, and the internal contradictions in the algorithm and the tables are exactly what a careful referee should catch. The good news is that the method idea is reasonable and the experiments are extensive enough to be worth salvaging. A major revision that fixes the update equations, clarifies the skip-communication branch, reconciles the ablations with the main tables, and releases code would make this a solid contribution to the FGL subfield.\n\nThis is a paper for FGL researchers and anyone thinking about communication-efficient personalization. It deserves a real review, not a desk rejection, because the core idea and experimental breadth justify referee time, even though the current form is not reproducible.","headline":"CEFGL's idea is plausible and its experiments are broad, but the algorithm as written is dimensionally invalid and inconsistent with itself; the headline numbers cannot be reproduced from the paper, so it needs major revision, not desk rejection.","tokens_in":661,"tokens_out":1281,"would_cite":false,"duration_ms":40894,"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":"CEFGL claims that decomposing each client's model into a low-rank shared component and a sparse personalized component simultaneously improves classification accuracy under non-IID graph data and cuts communication cost by an order of…","keywords":["Federated Graph Learning","Low-Rank Decomposition","Personalized Federated Learning","Sparse Models","Communication Efficiency","Non-IID Data","Quantization Compression"],"falsifier":"Re-run the CHEM cross-dataset experiment with GIN using only the published Algorithm 1 updates (Eqs. 6-9), the stated hyperparameters ($\\nu=0.5$, $\\alpha=0.6$, $\\mu=0.0001$, $\\lambda=0.001$, $p=0.5$), and 4-bit quantization, measuring accuracy and total transmitted bits over 200 rounds; if the accuracy gain over FedStar and the 18.58x bit reduction do not appear, the central claim fails. A simpler check is to verify that the aggregated global model after truncated SVD has rank near the claimed low-rank ratios in Table 3.","tokens_in":18567,"feed_emoji":"🔗","tokens_out":3682,"duration_ms":30790,"temperature":0.7,"pith_summary":"The paper proposes CEFGL, a federated graph learning algorithm aimed at two problems at once: clients' graph data are non-IID, so one shared model fits none of them well, and federated training is communication-bound. CEFGL decomposes each client model into a low-rank component that captures knowledge shared across clients and a sparse component that captures client-specific knowledge, trained through a dual-channel encoder. It adds a correction term to reduce client drift, allows many local updates between communication rounds, and compresses parameters by quantization. The paper reports that across sixteen datasets, under both IID and non-IID splits and with three GNN backbones, CEFGL achieves the best average classification accuracy compared with seven baselines. In the cross-dataset CHEM setting, it claims a 5.64% accuracy gain over FedStar while cutting communication bits by a factor of 18.58 and communication time by a factor of 1.65.","feed_headline":"Low-rank plus sparse models beat FedStar at 18x less communication","feed_subtitle":"CEFGL splits each client model into shared low-rank and private sparse parts, gaining 5.64% accuracy on CHEM with GIN.","key_machinery":"The central mechanism is parameter decomposition in the spirit of robust PCA: each client model is written as a low-rank component $W$ (global, shared) plus a sparse component $S$ (local, personalized), with convex relaxations of rank to trace norm and of $\\ell^0$ to $\\ell^1$. A dual-channel encoder trains $W$ and $S$ separately; a Scaffnew-style correction term $h_i^t$ counters client drift; local updates run many steps before communication; and a probabilistic skip in Algorithm 1 lets clients skip rounds. Quantization $Q_r$ compresses uplink and downlink messages to as few as 4 bits per parameter. The server performs truncated SVD aggregation, so the shared model stays low-rank.","core_discovery":"The central claim is that a federated graph classifier can be made both more accurate under data heterogeneity and dramatically cheaper to communicate by decomposing model parameters into a low-rank global part and a sparse personalized part, rather than transmitting dense full models. The paper argues that the low-rank component encodes common knowledge across clients, the sparse component encodes local knowledge, and the two together give each client a hybrid model that adapts locally without losing shared structure. It further claims the correction-term update and probabilistic communication skipping make the method robust to client dropout, while 4-bit quantization barely hurts accuracy. The headline numbers are stated against FedStar: +5.64% accuracy on the cross-dataset CHEM benchmark with GIN, 18.58x fewer communication bits, and 1.65x less communication time per round.","pith_inferences":["The same low-rank-global/sparse-local split is not graph-specific; it could be applied to federated image or text models, where the sparse component would play the role of a lightweight adapter.","The communication probability $p$ is effectively a tunable privacy-efficiency knob: lower $p$ means the server sees fewer updates and cannot reconstruct exact gradients, so the method could pair naturally with differential privacy.","A testable extension is to replace truncated SVD with a learned or adaptive rank selection per round, which might improve accuracy on datasets where the optimal rank varies across clients."],"forward_implications":["If CEFGL's claims hold, federated graph classification on heterogeneous cross-domain data can be made both more accurate and cheaper than state-of-the-art baselines, removing a major practical barrier to deployment.","The 18.58x reduction in communication bits means large GNN models can be trained federated on bandwidth-constrained or mobile clients without sacrificing accuracy.","The robustness to client dropout, with accuracy maintained even when the dropout rate is sampled from a Beta(10,1) distribution, implies the method suits real networks with unstable participation.","The finding that 10% of sparse parameters suffice for personalization suggests local personalization does not need a large private model, reducing on-device storage too."],"supporting_citations":[{"why":"Supplies the robust PCA decomposition $W+S$ that motivates splitting model parameters into low-rank and sparse parts.","marker":"[42]"},{"why":"Provides the correction-term (Scaffnew-style) update that reduces client drift and enables many local steps.","marker":"[33]"},{"why":"FedStar is the state-of-the-art FGL baseline that CEFGL compares against for the headline accuracy and communication numbers.","marker":"[40]"},{"why":"FedAvg is the standard federated learning baseline and the aggregation scheme CEFGL extends.","marker":"[19]"},{"why":"GCFL supplies the non-IID dataset partitioning and cross-dataset experimental setup used in the evaluations.","marker":"[24]"},{"why":"Supplies the quantization compression scheme $Q_r$ used to reduce uplink and downlink communication bits.","marker":"[46]"}],"fun_headline_variants":["Low-rank + sparse split cuts federated graph communication 18x","CEFGL: shared low-rank, private sparse, beats FedStar","Personalized federated graphs via low-rank global and sparse local","18x fewer bits: low-rank + sparse personalization for federated graphs","Federated graph learning: low-rank global, sparse local, 18x less comm"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's experimental numbers depend on the update equations in Section 4.1 and Algorithm 1 being the exact optimization steps that were run; if those equations do not match the implementation, the reported accuracy and communication gains cannot be reproduced.","fun_headline_variants_meta":{"raw":{"variants":["Low-rank + sparse split cuts federated graph communication 18x","CEFGL: shared low-rank, private sparse, beats FedStar","Personalized federated graphs via low-rank global and sparse local","18x fewer bits: low-rank + sparse personalization for federated graphs","Federated graph learning: low-rank global, sparse local, 18x less comm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001052,"raw_usage":{"total_tokens":4430,"prompt_tokens":967,"completion_tokens":3463,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":3365}},"tokens_in":583,"tokens_out":3463,"duration_ms":21799,"temperature":1.0,"reasoning_tokens":3365,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:07:55.739438+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the CHEM cross-dataset experiment with GIN using only the published Algorithm 1 updates (Eqs. 6-9), the stated hyperparameters ($\\nu=0.5$, $\\alpha=0.6$, $\\mu=0.0001$, $\\lambda=0.001$, $p=0.5$), and 4-bit quantization, measuring accuracy and total transmitted bits over 200 rounds; if the accuracy gain over FedStar and the 18.58x bit reduction do not appear, the central claim fails. A simpler check is to verify that the aggregated global model after truncated SVD has rank near the claimed low-rank ratios in Table 3.","supporting_citations":[{"cited_title":"Mishchenko, G","cited_arxiv_id":null,"evidence_quote":"Provides the correction-term (Scaffnew-style) update that reduces client drift and enables many local steps."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FedStar is the state-of-the-art FGL baseline that CEFGL compares against for the headline accuracy and communication numbers."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GCFL supplies the non-IID dataset partitioning and cross-dataset experimental setup used in the evaluations."},{"cited_title":"Alistarh, D","cited_arxiv_id":null,"evidence_quote":"Supplies the quantization compression scheme $Q_r$ used to reduce uplink and downlink communication bits."}],"review_version":1}