{"id":"5bdd9868-6d0d-406e-a0fb-5353fe43c981","arxiv_id":"2411.18919","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"POWER is a federated continual graph learning framework that combines replay of carefully selected old nodes with server-side prototype reconstruction and trajectory-weighted knowledge transfer, outperforming adapted baselines on eight graph datasets.","lead":"This paper introduces federated continual graph learning, where distributed clients each own a private graph that grows over time and a server combines their GNN training without sharing raw data. It proposes POWER, a framework using selected old-node replay and prototype-based knowledge transfer, and reports improvements in accuracy and forgetting across eight datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Absent significance testing, the Table 2 AM gains are within noise on most datasets; the 'consistently outperforms' central claim is not yet established.","rationale":"The reader's verdict is CONDITIONAL, and I agree with that outcome, but I identify statistical significance as the single most load-bearing concern rather than pseudo-prototype fidelity. The paper's central assertion is an empirical superiority claim across eight datasets. That claim is only as strong as the evidence in Table 2, and the reported means and standard deviations leave open the possibility that most AM differences are not significant. The pseudo-prototype reconstruction fidelity is a valid secondary concern, but the GEC module contributes modestly in the ablation (4% AM, 14% FM), and even a flawed GEC mechanism would not destroy the observed directional advantage if it were statistically robust. By contrast, if the AM gains evaporate under proper significance testing, the paper's headline result and its 'state-of-the-art performance' contribution would be unsupported. I therefore make significance the primary stress point while acknowledging the reader's additional concern. The proposed concrete test is directly executable because the authors have released source code, making per-seed reruns or analysis possible; this aligns with the paper's own reproducibility commitment and would settle the concern without new data collection.","tokens_in":23655,"tokens_out":8190,"duration_ms":77267,"concrete_test":"Using the released code, obtain per-seed AM and FM values for POWER and the strongest baselines (Fed-ERGNN, Fed-DSLR, and the best federated-CV method) on all eight datasets. Run paired Wilcoxon signed-rank tests across the 10 seeds for each dataset, and a two-sided binomial test on the eight dataset-level win/loss signs. Report the number of datasets where POWER is significantly better at p<0.05 for AM and FM separately. If fewer than six of eight datasets show significant AM gains, the 'consistently outperforms' claim in the abstract and Section 5.2 is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that POWER consistently outperforms all baselines in both AM and FM rests entirely on Table 2, which reports only means and standard deviations over 10 runs. No significance tests, confidence intervals, or paired comparisons are provided. For AM, the primary adaptation metric, the gap between POWER and the strongest federated CGL baseline is smaller than the combined standard errors on most datasets: Cora (65.74±5.11 vs 61.21±3.25), CiteSeer (54.47±5.21 vs 49.92±2.02), OGB-arxiv (36.19±3.17 vs 32.63±3.48), Computers (43.26±2.94 vs 40.32±5.30), Flickr (35.79±4.15 vs 31.55±2.47), and Roman-empire (26.54±2.35 vs 24.71±0.87). Although FM differences are larger and the directional consistency across eight datasets is suggestive, the claim of universal superiority is not statistically supported by the reported evidence. This concern is more load-bearing than pseudo-prototype fidelity because the empirical win is the paper's headline: if the wins are within noise, the central claim fails even if every module works as described.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Federated Continual Graph Learning (FCGL), a new problem setting in which multiple clients hold private evolving graphs and a central server coordinates GNN training under storage and privacy constraints. Through an empirical study on a Cora-based simulation, the authors identify two challenges: local graph forgetting (LGF), where a client loses prior knowledge when adapting to a new task, and global expertise conflict (GEC), where parameter aggregation across clients with divergent evolution trajectories makes the global GNN suboptimal on both past and current tasks. To address these, the paper proposes POWER, which (1) selects and replays experience nodes with maximum local-global coverage to mitigate LGF, and (2) reconstructs pseudo-prototypes from client gradient information and performs trajectory-aware knowledge transfer on a server-side global buffer graph to mitigate GEC. The method is evaluated on eight graph datasets against ten baselines, with additional ablations, sensitivity analyses, and sparsity experiments.","tokens_in":23890,"tokens_out":7123,"duration_ms":65849,"significance":"If the empirical results are taken at face value, the paper makes a useful contribution to the emerging area of federated continual graph learning. It is the first systematic treatment of the problem, with a wide benchmark (eight datasets across six domains), ten adapted baselines, an open-source implementation, and extensive secondary experiments (ablations, hyperparameter sensitivity, convergence, communication cost, and sparsity robustness). The LGF/GEC decomposition is intuitively grounded in the motivating empirical study, and the proposed two-module remedy is a reasonable first solution. The release of code and a technical report supports reproducibility. The main weaknesses are statistical: the headline claim of consistent superiority is not backed by significance testing, and the central pseudo-prototype mechanism is not directly validated for fidelity. These issues are fixable within the manuscript's scope rather than fatal.","major_comments":[{"comment":"The central claim that POWER \"consistently outperforms all baselines in both AM and FM\" is based solely on means and standard deviations over ten runs, with no significance tests, confidence intervals, or paired comparisons. Several AM gaps are modest relative to the reported variance, for example Computers (43.26±2.94 vs. Fed-DSLR 40.32±5.30) and CiteSeer (54.47±5.21, where the reported standard deviation is larger than the gap to the best baseline). With eight datasets and ten baselines, the multiple-comparison issue is nontrivial. I recommend adding paired significance tests (e.g., Wilcoxon signed-rank over datasets or paired bootstrap over runs) and reporting effect sizes or confidence intervals for the AM and FM differences. The directional consistency across all eight datasets is suggestive but does not by itself establish the claim.","section":"Sec. 5.2, Table 2"},{"comment":"The pseudo-prototype reconstruction mechanism is never validated for fidelity. The only evidence for the GEC module is the full-module ablation in Table 3 (\"w/o GEC\"), which removes both pseudo-prototype reconstruction and trajectory-aware knowledge transfer together; the intra-module ablation only swaps the cumulative trajectory for a non-cumulative one. Because the global buffer graph in Eq. (13) and the knowledge-transfer loss in Eq. (14) are built on these reconstructed pseudo-prototypes, noisy reconstructions would propagate directly into the server-side training signal. Please add a direct fidelity check (e.g., cosine similarity or nearest-class accuracy between reconstructed pseudo-prototypes and true class-mean prototypes) and a sensitivity analysis over reconstruction hyperparameters (LBFGS iterations, random seeds), or an ablation that replaces pseudo-prototypes with oracle prototypes. Without this, the mechanism's contribution to GEC is not isolated.","section":"Sec. 4.2, Eqs. (8)-(11)"},{"comment":"The privacy-preservation claim for prototype gradients is not supported. The text asserts that gradients from a randomly initialized network make it hard to extract meaningful information, but the paper itself shows that class labels are recoverable from the gradient sign (c = argmin_i ...), and gradient inversion is known to leak input information even through random networks. No formal privacy analysis, differential-privacy guarantee, or empirical attack evaluation is provided. Since privacy is one of the two motivating constraints of FCGL, this claim should be either substantiated or substantially qualified.","section":"Sec. 4.2, 'Pseudo Prototype Reconstruction Mechanism'"}],"minor_comments":[{"comment":"The hyperparameter list in A.4 refers to a threshold distance η, while Eq. (3) defines the same quantity as ε; the notation should be unified.","section":"A.4 vs. Eq. (3)"},{"comment":"Eq. (5) uses a binary cross-entropy form with a (1-y_i) log(1-ŷ_i) term for what is described as a multi-class one-hot label. If the classifier output is a softmax over classes, this loss is nonstandard; please clarify the intended output activation and loss formulation.","section":"Eq. (5)"},{"comment":"In the FedSage+ row for Squirrel, the AM and FM columns report identical values (29.52±2.18 and 79.82±3.04); this looks like a typo and should be checked.","section":"Table 2"},{"comment":"The notation p_k^i ∝ { sum ... } is ambiguous: please state explicitly whether p_k^i is normalized over classes or over samples, since the normalization affects the interpretation of the decayed cumulative trajectory q_t^k.","section":"Eq. (12)"},{"comment":"The sentence \"a small φ may fail to capture it effectively\" is confusing in light of Eq. (12), where smaller φ causes earlier tasks to decay more quickly; please rephrase to describe the effect of φ on the trajectory encoding.","section":"Sec. 5.4, Fig. 5"},{"comment":"The text contains a grammatical typo: \"an novel and effective FCGL training framework\" should read \"a novel and effective FCGL training framework.\"","section":"Sec. 1, Contributions"},{"comment":"Line 36 of Algorithm 1 says \"Reconstruct P̂ via ∇Γ_k^{t,c}\" but the details of the LBFGS optimization and its hyperparameters appear only in A.4; a brief pointer in the main text would improve reproducibility.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is empirical and the two self-citations [22,65] are background references; the central performance claims rest on external baselines, so there is no circularity concern. The main risk is that the KDD audience will find the statistical support insufficient: the paper should add significance tests or explicitly soften the 'consistently outperforms' claim. The pseudo-prototype mechanism also needs a direct sanity check, since the current ablation does not separate reconstruction fidelity from the other components of the GEC module. With those additions, the paper could be a solid contribution to a new problem area."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing you should know: this is a genuine first formulation of federated continual graph learning, with a coherent framework and a broad empirical study. The core claim—POWER beats federated adaptations of existing CGL methods—is probably right, but the paper would be stronger with significance tests and a direct validation of its pseudo-prototype module.\n\nWhat's new: the problem definition itself, the two named challenges (LGF and GEC), and the POWER framework. Local-global coverage replay is a sensible extension of experience replay to the federated setting. Pseudo-prototype reconstruction with trajectory-aware knowledge transfer is original, though it is the least-validated part. On experiments, the paper does a lot: eight datasets, ten baselines, ablations, sensitivity, sparsity, efficiency, and public code. That is solid, reproducible work.\n\nSoft spots, in order of real weight. First, no significance tests are reported. But the stress-test worry that the AM gains are within noise doesn't hold up if you compute standard errors from the reported standard deviations: on Cora, CiteSeer, OGB-arxiv, Flickr, and Roman-empire, the AM gap between POWER and the strongest federated CGL baseline is about two to three standard errors. FM gaps are large everywhere. So the central claim is defensible. The weakest case is Computers, and the absence of paired comparisons still makes 'consistently outperforms' too strong. Add significance tests and temper the wording.\n\nSecond, the GEC module. Pseudo-prototype reconstruction is a core advertised contribution, but the paper never checks whether the reconstructed pseudo-prototypes resemble true class prototypes. The only evidence is a single Cora ablation: removing GEC costs about 4 points AM and 14 points FM. That is not nothing, but it is thin for a mechanism that also carries a privacy claim. The privacy claim is itself under-motivated: sending gradients from a randomly initialized network is not an established privacy guarantee, and the class label is explicitly recoverable from the gradient. I would want either a reconstruction-fidelity analysis or a more careful privacy argument.\n\nCitation pattern is fine; the self-citations are background FGL and a technical report. Nothing circular or load-bearing.\n\nWho is this for? Anyone in federated graph learning or continual graph learning. It is a solid first step into a new intersection, not a paradigm shift. It deserves a serious referee; a revision should add significance testing and validate the pseudo-prototype module.","headline":"A real first formulation of federated continual graph learning with a broadly evaluated framework; the central claim mostly survives the noise objection, but the pseudo-prototype module needs stronger validation.","tokens_in":24416,"tokens_out":4684,"would_cite":true,"duration_ms":40772,"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":"POWER, the first federated continual graph learning framework, claims to beat all baselines on accuracy and forgetting across eight graph datasets by pairing local coverage replay with server-side pseudo-prototype transfer.","keywords":["federated continual graph learning","graph neural networks","catastrophic forgetting","experience replay","pseudo-prototype reconstruction","knowledge transfer","class-incremental learning","federated learning"],"falsifier":"Measure the cosine similarity between each reconstructed pseudo-prototype and the true class feature centroid across the eight datasets; low similarity would falsify the claim that the global buffer graph carries genuine class expertise rather than noise from the random encoding network.","tokens_in":23409,"feed_emoji":"🕸️","tokens_out":7972,"duration_ms":68077,"temperature":0.7,"pith_summary":"Until now, continual graph learning assumed one institution holds the whole evolving graph. This paper argues that real graph data lives across private databases, and proposes federated continual graph learning (FCGL) as the setting where multiple clients train GNNs on their own evolving graphs while a server aggregates their models. The paper reports that even naive federated training beats isolated continual graph learning, but two problems limit it: local graph forgetting within each client, and global expertise conflict, where averaging expert models leaves a global model worse than the best local expert. It then introduces POWER, which replays carefully selected experience nodes locally and performs trajectory-aware knowledge transfer on the server via reconstructed pseudo-prototypes. On eight datasets, POWER claims to consistently beat federated versions of existing CGL methods and vision-centric federated continual learning baselines on both accuracy and forgetting metrics.","feed_headline":"POWER beats ten baselines at federated continual graph learning","feed_subtitle":"Local replay and server-side prototype transfer cut forgetting on eight graph benchmarks.","key_machinery":"The argument runs on three coupled mechanisms. First, local-global coverage maximization picks experience nodes in embedding space using a weighted blend of local and global GNN embeddings, so replay samples represent a class from both one client's view and the federated view. Second, pseudo-prototype reconstruction lets the server learn class prototypes without raw features: clients upload gradients of class prototypes through a fixed randomly initialized encoding network, and the server recovers pseudo-prototypes by gradient matching, then builds a global buffer graph via K-nearest neighbors. Third, trajectory-aware knowledge transfer uses each client's cumulative label distribution over tasks (a decaying sum that encodes how its expertise evolved) as weights in a KL-divergence loss that pulls the global model toward the local experts most responsible for each class. Together these mechanisms target the two failure modes: replay fights LGF at the client, prototype transfer fights GEC at the server.","core_discovery":"The central discovery is that federated continual graph learning is not only feasible but necessary: with clients following divergent graph evolution trajectories, even a plain FedAvg combination of a continual graph learner outperforms isolated training, because different clients become experts in different classes at different times. The paper identifies two obstacles—local graph forgetting (LGF), the usual catastrophic forgetting within a client, and global expertise conflict (GEC), the loss of expertise that happens when parameters from differently specialized clients are averaged—and claims POWER resolves both. LGF is addressed by storing one experience node per class per task chosen to maximize combined local-global coverage in embedding space, then replaying those nodes during later tasks. GEC is addressed server-side: clients send gradient patterns of class prototypes rather than raw data, the server reconstructs pseudo-prototypes by gradient matching, builds a KNN graph over them, and uses each client's decaying cumulative label distribution (its evolution trajectory) to weight a KL-divergence distillation from local experts into the global model. The reported result is that POWER consistently beats the ten baselines in average accuracy and in forgetting across eight graph datasets spanning six domains.","pith_inferences":["A direct audit of the reconstruction step—comparing reconstructed pseudo-prototypes against true class centroids in feature space—would tell whether the server's global buffer graph carries genuine class expertise; the paper does not perform this check.","The trajectory encoding is label-only: representing client trajectories by cumulative label histograms ignores topology changes, so a natural extension would incorporate graph-structure drift (edge additions, community shifts) into the trajectory weights.","The local-global coverage selector could be lifted out of the continual setting and used as a privacy-aware buffer construction for any federated graph learner, since it needs only embeddings from local and global models.","The KL-transfer step is server-side, so it should also benefit asynchronous or partially participating clients whose prototypes were stored in earlier rounds, a property the sparsity experiments support but do not fully stress-test."],"forward_implications":["Federated deployment of continual graph learning is viable: the paper's experiments indicate that collective intelligence across clients outweighs isolated training even before any new algorithmic machinery is added.","Experience replay for graphs can be made federation-aware by choosing buffer nodes with both local and global embeddings; this yields better retention than selecting with local embeddings alone, as the ablation on Cora shows.","Raw class prototypes need not leave the client: gradient matching from a randomly initialized network supplies the server with usable class knowledge, keeping the raw-feature privacy boundary intact.","The extra communication cost of the GEC module is one small prototype-gradient upload per task, compared with GNN parameter uploads every round, so conflict resolution is cheap relative to standard federated communication.","The framework stays effective under missing features, missing edges, missing labels, and partial client participation, which are exactly the conditions decentralized evolving graphs present."],"supporting_citations":[{"why":"supplies the coverage-maximization experience replay idea that POWER extends with global embeddings to select local buffer nodes.","marker":"[63]"},{"why":"provides the TWP regularization baseline that is evaluated in federated form as Fed-TWP.","marker":"[25]"},{"why":"provides the DSLR diversity-based replay baseline evaluated as Fed-DSLR.","marker":"[7]"},{"why":"defines FedAvg, the parameter-aggregation backbone for all federated variants in the paper, including POWER itself.","marker":"[30]"},{"why":"gives the FedSage+ subgraph federated learning baseline in the FL/FGL fine-tuning comparison.","marker":"[59]"},{"why":"gives the FedGTA topology-aware averaging baseline in the FL/FGL fine-tuning comparison.","marker":"[22]"},{"why":"gives the GLFC vision-oriented federated continual learning baseline that POWER is compared against.","marker":"[9]"},{"why":"gives the TARGET exemplar-free distillation baseline from federated class-continual learning for comparison.","marker":"[57]"},{"why":"gives the LANDER text-anchored data-free baseline from federated continual learning for comparison.","marker":"[41]"},{"why":"supplies the community detection method used to partition datasets into client subgraphs for the simulation.","marker":"[5]"}],"fun_headline_variants":["POWER fixes local forgetting and global conflict in federated graphs","Forgetting less, learning more: POWER on federated graph streams","POWER beats ten baselines on eight graph benchmarks","Decentralized graph learning: POWER turns local experts into global gains","POWER tames catastrophic forgetting in distributed evolving graphs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that pseudo-prototypes reconstructed by gradient matching from a randomly initialized network are faithful stand-ins for each client's class expertise; if the reconstruction is noisy, the server-side buffer graph and knowledge transfer will spread distorted expertise.","fun_headline_variants_meta":{"raw":{"variants":["POWER fixes local forgetting and global conflict in federated graphs","Forgetting less, learning more: POWER on federated graph streams","POWER beats ten baselines on eight graph benchmarks","Decentralized graph learning: POWER turns local experts into global gains","POWER tames catastrophic forgetting in distributed evolving graphs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001012,"raw_usage":{"total_tokens":4287,"prompt_tokens":966,"completion_tokens":3321,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":3237}},"tokens_in":582,"tokens_out":3321,"duration_ms":21772,"temperature":1.0,"reasoning_tokens":3237,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:44:12.683613+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the cosine similarity between each reconstructed pseudo-prototype and the true class feature centroid across the eight datasets; low similarity would falsify the claim that the global buffer graph carries genuine class expertise rather than noise from the random encoding network.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the coverage-maximization experience replay idea that POWER extends with global embeddings to select local buffer nodes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the DSLR diversity-based replay baseline evaluated as Fed-DSLR."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"gives the FedSage+ subgraph federated learning baseline in the FL/FGL fine-tuning comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"gives the FedGTA topology-aware averaging baseline in the FL/FGL fine-tuning comparison."},{"cited_title":"Cohen, and Ruslan Salakhutdinov","cited_arxiv_id":null,"evidence_quote":"gives the TARGET exemplar-free distillation baseline from federated class-continual learning for comparison."}],"review_version":1}