{"id":"e0a644f1-b87d-41ca-851e-46c35bc1c002","arxiv_id":"2411.12098","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"FCLG combines local graph augmentation contrastive learning with model-level contrastive distillation to learn unsupervised graph embeddings in a federated setting, beating InfoGraph and MVGRL baselines on graph clustering.","lead":"This paper proposes FCLG, a federated learning method that learns graph-level representations without labels by applying contrastive learning both inside each local data silo and between local and global models. On four public benchmark datasets, FCLG reports clustering accuracy gains of 2 to 7 percent over federated versions of existing graph embedding methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline margins rest on per-dataset test-label hyperparameter tuning and an appendix that is absent; the claimed 2-7% gains are not yet shown to be robust.","rationale":"The reader's weakest assumption identifies test-label hyperparameter tuning as the load-bearing risk, and the full text supports that reading: Section IV-A explicitly tunes all hyperparameters for optimal clustering performance on each dataset, and Section V-A does the same for every point in Figure 4. I agree this is the single most serious threat to the paper's central empirical claim. I mark agreement as 'partial' because I would sharpen the concern in two ways: the missing appendix and code make the tuning protocol impossible to audit, and the paper's own variant analysis shows that removing inter-contrasting (Intra-FedAvg) matches or slightly beats FCLG on PROTEINS and NCI1 in both IID and non-IID tables, so the specific contribution of the inter-level contrast is not uniformly supported. That said, the overall claim that the FCLG framework outperforms the constructed baselines does not depend on every component helping on every dataset. No internal inconsistency was found in the derivations; Eqs. 5-8 are standard contrastive objectives, and Algorithm 1 is coherent. There is no formal verification or released code, so independent support is limited. A fixed-protocol rerun is the one check that would either rescue the headline margins or show they were an artifact of per-dataset test-label selection; hence I do not move the verdict away from CONDITIONAL.","tokens_in":16005,"tokens_out":5228,"duration_ms":62201,"concrete_test":"Release the missing appendix and code; fix one hyperparameter configuration per method chosen without test labels (for example, selected by silhouette score on client-local embeddings or on a held-out labeled split), rerun Table II with identical client partitions and seeds, and report mean plus/minus standard deviation with the margin over the closest baseline. If under this protocol FCLG's average advantage is below 2% on more than one dataset, the 'significant margins' claim should be softened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—a 'significant clustering gain' of 2-7% over the closest baseline—requires that the comparison protocol does not select models on the evaluation labels. Section IV-A states 'all hyper-parameters tuned for optimal clustering performance on each dataset', and Section V-A states that the points in Figure 4 are obtained 'by fine-tuning hyper-parameters separately'. Because clustering accuracy and macro F1 on the test-set labels are the headline metrics, tuning on those labels means the reported margins are upper envelopes over hyperparameter choices rather than the performance of a fixed, defensible configuration. The paper also says detailed hyperparameter settings are in the Appendix, but the appendix is not present in the manuscript, and no code is released, so the tuning grid and its cost are unverifiable. FCLG introduces additional hyperparameters (intra temperature tau, inter temperature tau-prime, local training epochs, and the linear combination in Eq. 9) beyond the baselines; if FCLG's grid was searched more aggressively, some or all of the 2-7% gap could be a tuning artifact. The concern is not that the tables are fabricated; it is that the central empirical claim is conditional on an uncontrolled selection procedure.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes FCLG, a federated learning framework for unsupervised graph-level representation learning. FCLG combines two contrastive objectives: an intra-client instance-wise contrastive loss over augmented views of graphs (adapted from SimCLR/MVGRL-style augmentation) and an inter-client model-level contrastive loss that aligns the current local model's graph-level representations with the global model's representations while repelling the previous local epoch's representations (adapted from MOON). The authors evaluate the learned embeddings via K-Means clustering on four TUDatasets (PROTEINS, ENZYMES, DHFR, NCI1) in both IID and non-IID federated settings, comparing against InfoGraph and MVGRL combined with FedAvg and FedProx. They report 2-7% clustering accuracy gains over the closest baselines in non-IID settings and 2-4% in IID settings, alongside ablations and robustness experiments.","tokens_in":16376,"tokens_out":8095,"duration_ms":78499,"significance":"The contribution is potentially significant: if the empirical claims hold, this is an early demonstration that unsupervised graph-level representation learning can be federated without labels, and the two-level contrastive design is a sensible combination of existing ideas. The paper's strengths include the clear algorithmic description, the use of 10-run repeats for main tables, and the ablation of inter-contrast variants (Intra-FedAvg, Intra-KL, Intra-MSE). However, the empirical validation has a load-bearing flaw: hyperparameters are tuned to maximize the evaluation metric on the same dataset labels used for the reported results, and the detailed settings are relegated to a missing appendix. The 2-7% margins can therefore not currently be attributed to the method rather than to the tuning protocol. The absence of significance tests and error bars further weakens the 'significant gain' claim.","major_comments":[{"comment":"The central empirical claim is undermined by the hyperparameter selection protocol. Section IV-A states that 'all hyper-parameters are tuned for optimal clustering performance on each dataset,' and the caption of Figure 4a/4b says each result is 'obtained by fine-tuning hyper-parameters separately.' Because the metric used for tuning (clustering accuracy and macro F1 against ground-truth labels) is the same as the headline evaluation metric, the reported gains in Tables II and III (e.g., 7% on PROTEINS, ~5% on ENZYMES and DHFR, 2% on NCI1) are upper envelopes over the hyperparameter grid rather than the performance of a defensible fixed configuration. To support the claim of 'significant clustering gain,' the authors should either fix hyperparameters in advance, select them on a validation set without using the evaluation labels, or show that the reported margins are stable across the tuning grid.","section":"IV-A and V-A"},{"comment":"The manuscript is not reproducible as submitted. The text refers twice to an Appendix for 'detailed hyper-parameter settings' and for the official baseline code, but no appendix is included in the manuscript. This is particularly problematic because FCLG introduces additional hyperparameters (τ, τ′, the number of local epochs E, and the loss combination in Eq. 9) beyond the baselines. Without the actual grid, the reader cannot assess whether FCLG and the baselines received comparable tuning budgets, which is exactly the point on which the headline comparison rests. The authors should include the appendix and release code (or at least the full configuration and search ranges) before the results can be verified.","section":"IV-A"},{"comment":"The claim of a 'significant clustering gain' is not backed by statistical inference. The 10-run ranges in Tables II and III show that the mean differences are large relative to the spread, but no significance test is reported, and the ranges are conditional on the per-dataset hyperparameter selection discussed above. In addition, Figures 2, 4, and 5 present trajectories and comparisons without any error bands, so the robustness of the skew (Fig. 4a) and client-number (Fig. 4b) experiments cannot be assessed. The authors should add significance tests and error bars/bands, or at least report per-run values for the experiments in Figures 4 and 5.","section":"IV-B"}],"minor_comments":[{"comment":"The row label 'FGCL' should be 'FCLG' to match the method name used throughout the paper.","section":"Table III"},{"comment":"The caption says 'on PROTEINS (20 communication rounds)' but the figure contains four lines labeled PROTEINS, ENZYMES, DHFR, and NCI1; the caption should either list all datasets or state that all four are shown.","section":"Figure 5 caption"},{"comment":"The text describing FCLG-H says 'intra-contrasting is conducted between node-level representations H via Eq. 8,' but Eq. 8 defines an inter-contrasting variant (l^H_inter); the text should say 'inter-contrasting' rather than 'intra-contrasting.'","section":"Section IV-B"},{"comment":"The contrastive loss in Eq. (5) is written as a positive log ratio; to align with the standard InfoNCE formulation and the statement that the objective is minimized, it would be clearer to write it with an explicit negative sign or to state explicitly that minimizing L maximizes agreement between positive pairs.","section":"Eq. (5)"},{"comment":"The procedure for creating the non-IID client splits is not described in the main text (e.g., whether a Dirichlet allocation over classes is used); since the non-IID setting is central to the evaluation, this detail should be provided in the main text or in the appendix.","section":"Section IV-A"},{"comment":"The clustering evaluation protocol does not state how the K-Means cluster labels are matched to ground-truth classes (e.g., Hungarian matching); this step should be described for reproducibility.","section":"Section IV-B"}],"recommendation":"major_revision","confidential_remarks":"The core idea is reasonable and the paper is clearly written, but the empirical evaluation needs substantial revision. The editor may want to require the authors to provide the missing appendix, release code, and a fair hyperparameter selection protocol that does not use the evaluation labels. If the authors cannot provide such a protocol, the central claims should be softened accordingly. Also, given the novelty claim, the authors should carefully check for prior work on federated graph-level representation learning that may have appeared since this submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"FCLG is a reasonable first pass at unsupervised federated graph-level representation learning. The new combination is intra-client augmentation contrast (SimCLR/MVGRL-style) plus inter-client model contrast (MOON-style) applied to graph-level embeddings. The paper claims 2-7% clustering gains over federated baselines on four TUDatasets; those gains are plausible, but the current evidence does not nail down their size.\n\nWhat the paper does well: The problem is clearly motivated and the method section is readable. The authors construct sensible baselines by combining InfoGraph and MVGRL with FedAvg and FedProx, which is the right thing to do in a setting where no direct prior work exists. The experiments cover non-IID and IID splits, include 10-run ranges, and ablate the inter-client loss against FedAvg-only, KL, and MSE variants. Those ablations support the design choice. The sensitivity analyses for EMD, client count, and local epochs are also useful.\n\nThe soft spots are real and mostly about the evaluation protocol. Section IV-A says all hyperparameters were tuned for optimal clustering performance on each dataset, and Section V-A says the points in Figure 4 were obtained by fine-tuning hyperparameters separately. Since clustering accuracy on the test labels is the headline metric, tuning on those labels turns the reported margins into upper envelopes over the hyperparameter grid. The appendix containing the actual settings is absent, and no code is released, so the cost and size of that grid are unverifiable. There are also no significance tests, and Figures 2, 4, and 5 lack error bars. None of this suggests fabrication, but it does mean the 2-7% claim could shrink under a fixed, preregistered configuration. Minor issues: Table III has a row labeled 'FGCL' that should be FCLG, and the stated DHFR/NCI1 margins don't exactly match Table II, but those are easy fixes.\n\nWho this is for: anyone working on federated graph learning or privacy-preserving representation learning. It is a legitimate first step in an underexplored subarea, and the method is clearly specified enough to reimplement. It deserves a serious referee, but only with the expectation of major revision. I would send it to review and ask for code, the full hyperparameter grid, a nested or fixed protocol for model selection, and significance tests. The core idea is worth keeping.","headline":"FCLG is a reasonable first pass at unsupervised federated graph-level representation learning, but the claimed 2-7% gains currently rest on a tuning protocol that selects hyperparameters on test labels, with the appendix and code missing.","tokens_in":16777,"tokens_out":2980,"would_cite":true,"duration_ms":31390,"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":"This paper claims that a two-level contrastive framework, FCLG, can learn unsupervised graph-level representations in federated settings and reports clustering gains of 2–7% over federated baselines.","keywords":["federated learning","graph-level representation learning","contrastive learning","unsupervised learning","graph clustering","non-IID data","graph neural networks","knowledge distillation"],"falsifier":"Rerun the non-IID comparison with a single fixed hyperparameter configuration chosen without access to test labels—either identical across datasets or selected by an unsupervised criterion—and check whether FCLG's 2–7% accuracy margins over InfoGraph-FedProx survive; if they shrink to within run-to-run variance, the superiority claim fails.","tokens_in":15803,"feed_emoji":"🧬","tokens_out":8907,"duration_ms":85974,"temperature":0.7,"pith_summary":"Graph-level representations—embeddings of entire graphs rather than individual nodes—are needed for tasks like predicting protein properties or flagging malicious network traffic, but the graphs often cannot be pooled into one place. This paper tries to establish that such representations can still be learned without labels when data are scattered across isolated clients, by applying contrastive learning at two levels. Inside each client, augmented views of the same graph are pulled together and other graphs are pushed apart; between client and server, the current local representation is pulled toward the global model and away from the local model's previous round, which the paper argues counters non-IID drift. On four public protein and molecule benchmarks, FCLG reports 2–7% clustering accuracy gains over federated baselines in non-IID splits and 2–4% gains in IID splits. If this holds, institutions with private graph data could jointly train reusable graph embeddings without sharing raw graphs.","feed_headline":"Two-level contrastive learning beats federated graph baselines by 2–7%","feed_subtitle":"Unsupervised embeddings trained across isolated sites improve graph clustering on four protein and molecule benchmarks.","key_machinery":"The load-bearing mechanism is the two-level contrastive loss. Intra-contrasting generates a second view of each input graph through graph diffusion, treats that view as the sole positive pair, and uses all other graphs in the batch as negatives, following the SimCLR objective, which forces the local encoder to produce well-separated graph embeddings. Inter-contrasting compares three graph-level representations of the same batch: the current local model's output, the local model's output from the previous local epoch, and the global model's output, minimizing a contrastive loss that increases similarity between current-local and global while decreasing similarity between current-local and previous-local. This second term is the paper's answer to non-IID drift: it constrains each client's local updates to stay aligned with the aggregated global view while preventing the local model from wandering along its own skewed distribution.","core_discovery":"The paper's central claim is that a single two-level contrastive objective solves both the unsupervised-learning problem and the federated non-IID problem at once. The first level, intra-contrasting, runs instance-wise contrastive discrimination on graph-level representations inside each client, so each graph becomes distinguishable from all other graphs in the local batch. The second level, inter-contrasting, uses a cosine-similarity contrastive loss that pulls the current local representation toward the global model's representation and pushes it away from the representation produced by the local model in the previous local epoch. The paper argues this inter term is an advanced ensembling-distillation technique and shows that removing it costs about 5% on DHFR and 2% on ENZYMES, while replacing it with KL or MSE distillation costs 3–5% on PROTEINS and DHFR. Across four datasets, FCLG and its node-level variant FCLG-H outperform InfoGraph and MVGRL combined with FedAvg or FedProx, with the largest margins in non-IID settings.","pith_inferences":["Our inference: the same two-level recipe should transfer to other decentralized unsupervised settings—time series, point clouds, or text—because only the graph-diffusion augmentation is domain-specific, while both contrastive losses are domain-agnostic.","Our inference: the knowledge-distillation framing suggests a cheaper or stronger variant might replace the previous-local-round negative with a server-side prototype or a moving average of global representations; comparing those variants would isolate what the temporal negative contributes.","Our inference: because the four benchmarks are small and skewed by class label only, the open question is whether the margins survive on large federated graph collections where client drift is driven by topology rather than label proportions."],"forward_implications":["Private graph holders can jointly train an unsupervised graph encoder without sharing raw graphs, and the learned embeddings separate classes well enough to beat locally trained federated baselines in K-Means clustering.","The inter-contrastive term is doing real work: dropping it lowers accuracy by about 5% on DHFR and 2% on ENZYMES, so the margin is not solely from the intra-client contrastive loss.","The method's advantage grows as client distributions become more skewed, since the paper reports larger relative gains at higher EMD values.","Using node-level rather than graph-level representations for inter-contrasting gives similar clustering performance, so the benefit does not depend on the pooling granularity of the contrastive alignment."],"supporting_citations":[{"why":"Supplies the FedAvg aggregation rule that forms the global-model update and the baseline combinations.","marker":"[15]"},{"why":"Motivates the non-IID federated graph problem and the settings used to build the baselines.","marker":"[16]"},{"why":"Provides the InfoGraph unsupervised graph-level method that FCLG builds on and that serves as a baseline.","marker":"[19]"},{"why":"Supplies the GIN encoder and jumping-knowledge style concatenation used to compute graph representations.","marker":"[22]"},{"why":"Provides the MVGRL multi-view contrastive method used as a second baseline and the graph-diffusion augmentation idea.","marker":"[29]"},{"why":"Supplies the instance-wise contrastive objective and temperature parameter that intra-contrasting adapts.","marker":"[30]"},{"why":"Supplies the model-level contrastive mechanism that inter-contrasting extends from image models to graph models.","marker":"[32]"},{"why":"Provides the graph diffusion preprocessing used to generate the augmented view for intra-contrasting.","marker":"[51]"},{"why":"Supplies the four benchmark datasets on which the clustering evaluation is run.","marker":"[56]"},{"why":"Supplies the FedProx proximal-term federated optimizer used in baseline combinations.","marker":"[60]"}],"fun_headline_variants":["Federated graph clustering gets a dual contrastive boost","Two-level contrastive learning tames federated non-IID graphs","Contrastive learning at two levels lifts federated graph models","Dual contrastive objectives improve federated graph clustering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported margins rest on the assumption that tuning all methods' hyperparameters on each dataset's true labels is a fair comparison, and that four small protein and molecule benchmarks with class-proportion skew capture the federated graph-level setting.","fun_headline_variants_meta":{"raw":{"variants":["Federated graph clustering gets a dual contrastive boost","Two-level contrastive learning tames federated non-IID graphs","Contrastive learning at two levels lifts federated graph models","Dual contrastive objectives improve federated graph clustering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000624,"raw_usage":{"total_tokens":2908,"prompt_tokens":984,"completion_tokens":1924,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":1855}},"tokens_in":600,"tokens_out":1924,"duration_ms":14202,"temperature":1.0,"reasoning_tokens":1855,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:54:04.872083+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the non-IID comparison with a single fixed hyperparameter configuration chosen without access to test labels—either identical across datasets or selected by an unsupervised criterion—and check whether FCLG's 2–7% accuracy margins over InfoGraph-FedProx survive; if they shrink to within run-to-run variance, the superiority claim fails.","supporting_citations":[{"cited_title":"Communication-efficient learning of deep networks from decentralized data,","cited_arxiv_id":null,"evidence_quote":"Supplies the FedAvg aggregation rule that forms the global-model update and the baseline combinations."},{"cited_title":"Federated Graph Classification over Non-IID Graphs","cited_arxiv_id":"2106.13423","evidence_quote":"Motivates the non-IID federated graph problem and the settings used to build the baselines."},{"cited_title":"Infograph: Unsupervised and semi-supervised graph-level representation learning via mutual information maximization,","cited_arxiv_id":null,"evidence_quote":"Provides the InfoGraph unsupervised graph-level method that FCLG builds on and that serves as a baseline."},{"cited_title":"How powerful are graph neural networks?","cited_arxiv_id":null,"evidence_quote":"Supplies the GIN encoder and jumping-knowledge style concatenation used to compute graph representations."},{"cited_title":"Contrastive multi-view representa- tion learning on graphs,","cited_arxiv_id":null,"evidence_quote":"Provides the MVGRL multi-view contrastive method used as a second baseline and the graph-diffusion augmentation idea."}],"review_version":1}