{"id":"c21294a6-7ed3-4e3e-843b-b4513a8a5f39","arxiv_id":"2411.08374","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"FedGLS trains federated graph models when some clients have only node features and no graph edges, by distilling structure knowledge into a feature encoder and generating graphs locally on graphless clients.","lead":"Federated learning lets organizations train models together without sharing data, but some participants may lack the connections between data points. This paper introduces FedGLS, which lets those 'graphless' clients learn the missing connections from other clients and still train a graph neural network.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Global feature-encoder aggregation is the only transfer channel, but Eq. 11 overwrites it on graphless clients; no ablation isolates it, so the claimed structure transfer may not be the source of the gains.","rationale":"The reader's CONDITIONAL verdict is appropriate. I agree that the weakest assumption is that FedAvg of phi transfers usable graph-structure knowledge, and I also see the partial circularity the reader notes. My concern sharpens this into a specific, testable gap: the paper never ablates the phi transfer channel, so the experimental results cannot distinguish between genuine structure transfer and a local self-training effect. This is not an accusation of fraud and does not require rejecting the paper; it is an omission that a controlled experiment can settle. The proposed ablation is minimal and directly targets Eq. 12, Eq. 11, and Eq. 9. If the ablation preserves accuracy, the central claim is overstated; if it collapses accuracy, the claim is confirmed. The reader's other points (weak baselines, lack of significance tests) are valid but secondary; the mechanism question is the most load-bearing because the paper's novelty rests on it.","tokens_in":17887,"tokens_out":7006,"duration_ms":71257,"concrete_test":"On Cora and PubMed, rerun the Table 1 protocol with a single change: aggregate phi only from structured clients in Eq. 12; graphless clients receive the server's phi at the start of each round but do not update phi locally (skip Eq. 11 for all c in C2) and do not return phi to the server. Keep all other hyperparameters and the theta aggregation identical. Compare the average accuracy on graphless clients against the full FedGLS numbers across the same five seeds. If the accuracy stays within one standard deviation, then cross-client phi transfer is not the operative mechanism; if it drops substantially, the transfer channel is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central mechanism is that the globally aggregated feature encoder phi (Eq. 12) carries structure knowledge from structured clients to graphless clients, and that the graph learner is trained by Eq. 9 to fit that knowledge. However, on every graphless client, phi is updated during the same local round by knowledge distillation (Eq. 11) to mimic the local GNN, whose input adjacency S is produced by the graph learner itself. Thus, after the first round, the target H in Eq. 9 is not an independent cross-client teacher: it is a blend of a global signal and a self-referential copy of the local GNN output on the locally generated graph. The paper reports no ablation that isolates the transfer channel. The observed accuracy gains over Fed-GNNk in Table 1 could therefore be produced by the graph learner fitting local feature and label structure through the CE-trained theta (Eq. 10), rather than by structure knowledge transferred from other clients. If the phi channel is inert, the central claim that graphless clients learn from other clients' structure is unsupported, even if the accuracy numbers are reproducible.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces FedGLS, a federated graph learning framework for settings where some clients have complete graph data (node features plus edges) while other 'graphless' clients have only node features. Each client trains a GNN and a feature encoder; on graphless clients, a learnable graph learner generates a local adjacency matrix, and the feature encoder is trained by knowledge distillation from the GNN. Structure knowledge is claimed to be transferred across clients by FedAvg aggregation of both the GNN parameters and the feature-encoder parameters. The graph learner is optimized with a contrastive loss that encourages the GNN embeddings on the generated graph to agree with feature-encoder embeddings. Experiments on Cora, CiteSeer, PubMed, Flickr, and ogbn-arxiv compare FedGLS against five baselines and an oracle that uses true graph structures, with sensitivity studies over local epochs and graphless-client ratios. The paper reports that FedGLS consistently achieves the highest accuracy among the five baselines and converges faster than Fed-GNNk, the best baseline.","tokens_in":18121,"tokens_out":5258,"duration_ms":54105,"significance":"If the structure-transfer mechanism is sound, the paper addresses a novel and practically relevant gap: many federated systems have clients that lack explicit edge information, and the proposed pipeline offers a way for such clients to benefit from structure knowledge held by other clients. The formal problem definition, the clear algorithmic description, and the breadth of experiments across five datasets are strengths, as are the sensitivity studies over local epochs and graphless-client ratios. The manuscript also states that an implementation is available. However, the central claim that graphless clients learn structure from other clients’ knowledge is not isolated experimentally, and the distillation loop in Eqs. (9)–(11) creates a self-referential training signal that could allow the graph learner to succeed without any cross-client structure transfer. Because the paper's contribution is precisely that transfer channel, the current evidence falls short of establishing the mechanism, even if the reported accuracy numbers are reproducible.","major_comments":[{"comment":"The claimed structure-transfer mechanism is not isolated. On every graphless client, phi is updated by the KD loss in Eq. (11), where the teacher is the local GNN whose input adjacency S is produced by the graph learner itself. After the first round, the global phi aggregated in Eq. (12) therefore contains a self-referential component from graphless clients, and the H used as the target in the contrastive loss Eq. (9) is not an independent cross-client teacher. The paper reports no ablation that removes or varies this transfer channel, such as freezing phi on graphless clients, excluding graphless-client phi from aggregation, or using a teacher derived only from structured clients. Without such an ablation, the accuracy gains over Fed-GNNk in Table 1 could be produced by the graph learner fitting local label structure through the CE-trained GNN (Eq. 10), rather than by structure knowledge transferred from other clients. This is load-bearing for the paper's central claim.","section":"Sec. 4.2, Eq. (11); Algorithm 1"},{"comment":"The baseline set is too weak to support the claim that federated structure transfer is the source of the gains. Fed-GNNk and Local-GNNk use fixed kNN graphs, which are known to be poorly suited to node classification. There is no baseline that applies a learnable graph-structure-learning method (e.g., an adapted IDGL or SL-style learner) to graphless clients, and no ablation in which the graph learner is trained without any cross-client aggregation. Such a comparison is needed to distinguish the benefit of transferred structure knowledge from the benefit of adaptively learning a task-specific graph from local features and labels.","section":"Sec. 5.1.2, Table 1"},{"comment":"No statistical significance tests are reported, despite several small margins relative to the reported standard deviations. For example, on PubMed FedGLS reports 0.8491±0.0070 versus Fed-GNNk's 0.8426±0.0128, and on CiteSeer FedGLS reports 0.8058±0.0171 versus FedProto's 0.7876±0.0184. The claim that FedGLS 'consistently achieves the highest classification accuracy' requires paired significance tests or confidence intervals over the five repetitions, especially because the differences are not uniformly large.","section":"Table 1"},{"comment":"The convergence-speed claim is supported only on Cora and Flickr, and the curves in the submitted text are not readable. Please report convergence curves for all five datasets, or justify why the two selected datasets are representative, and state clearly whether the plotted loss is the average local training loss or the global objective. This matters because the claim that the adaptive graph learner accelerates convergence is part of the paper's experimental contribution.","section":"Sec. 5.2.2, Fig. 3"}],"minor_comments":[{"comment":"The denominator in Eq. (8) is written ambiguously: the summation upper limit should be explicitly n(k), and the bracket placement after 1[i≠j] should be clarified to show that both exponentials are inside the sum.","section":"Eq. (8)"},{"comment":"The KL divergence uses f(z_i; theta_c) and f(h_i; theta_c), where theta_c is the shared classifier head, but the paper does not explicitly state that theta_c is updated jointly with the GNN encoder via Eq. (10) and is also part of the aggregated theta in Eq. (12). Please clarify the role and update rule of theta_c.","section":"Sec. 4.2, Eq. (11)"},{"comment":"The complexity analysis covers only the GCN and MLP forward/backward costs. It omits the graph learner, the adjacency generation, and the kNN sparsification cost, so the conclusion that the feature encoder adds negligible cost is not a complete comparison of FedGLS against Fed-GNNk.","section":"Appendix C"},{"comment":"The Fed-GNN column is listed in Table 1 but is not described as a baseline in the Experimental Setup section. Please state explicitly that it is an oracle or upper-bound method using true graph structures on graphless clients.","section":"Sec. 5.1.2, Table 1"},{"comment":"The graph learner is updated only once per round, before the E local epochs for theta and phi, while theta and phi are updated E times. This asymmetry is not discussed, and the sensitivity study only varies E; the effect of the number of graph-learner updates per round is not explored.","section":"Algorithm 1"},{"comment":"The notation C1 = {c(k)}_{k=1}^{M} and C2 = {c(k)}_{k=M+1}^{K} is introduced but not used consistently in later sections; for example, Sec. 4.4 refers to 'clients in C1' without restating the range of k. A table or explicit restatement of the two client sets would improve readability.","section":"Sec. 3.2, Problem 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript header indicates the paper was published in TMLR in 10/2024; if this is a post-publication review, the editor may wish to weigh whether the requested ablation is essential to the published claim or can be treated as future work. The stress-test concern about the self-referential distillation loop is the most serious issue: the central conceptual contribution is the transfer of structure knowledge via aggregated phi, but the current experimental design does not demonstrate that this channel is responsible for the observed gains. I recommend major revision rather than rejection because the proposed problem is novel and the framework is plausible, but the load-bearing evidence is missing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth knowing: this paper formalizes a real and previously unhandled setting in federated graph learning—some clients have node features only, no graph structure—and proposes FedGLS, a reasonably clean combination of a feature encoder distilled from a GNN, a local graph learner trained with a contrastive loss, and FedAvg-style aggregation of both θ and φ. The experiments on five datasets show mean accuracy improvements over five baselines, including a kNN-preprocessing baseline, and they include Fed-GNN with real structure as an upper bound. The problem statement is clear, the algorithm is well specified, and the authors acknowledge a genuine limitation: the learned graphs use fixed-k neighbors and may not match the true structure. The related work is well positioned, covering FGL methods and graph-less MLPs like GLNN.\n\nThe main soft spot is the transfer mechanism. The graph learner on a graphless client is trained to maximize consistency between GNN embeddings Z and feature-encoder embeddings H, but H is itself distilled from Z on the same client (Eq. 11). After local KD, φ is a mix of the global φ and a feature-only student of the local GNN. So the contrastive objective is not an independent signal from structured clients; it mostly aligns the graph learner with the local GNN's own output. The only external structure knowledge enters through global aggregation of θ and φ, and the paper reports no ablation isolating either channel. The stress-test note is fair on this point. It doesn't kill the paper, because global θ is also transferred and could drive the gains, but the central claim—that graphless clients learn structure from other clients via φ—is not actually demonstrated.\n\nSecond, the experimental evidence is weaker than the text suggests. No significance tests are reported, and on several datasets the gap between FedGLS and the best baseline is less than one standard deviation (e.g., CiteSeer: 0.8058±0.0171 vs 0.7884±0.0131; PubMed: 0.8491±0.0070 vs 0.8426±0.0128). With five runs, those differences could easily be noise. Also missing is a natural baseline: a graphless client learning its own graph structure locally without any federation. Without that, it's hard to tell how much of the gain comes from federation versus from the graph learner architecture itself.\n\nThese are addressable rather than fatal. The problem is new, the method is sensible, the writing is honest, and the results are at least suggestive. Who is this for: anyone working on federated graph learning, and to a lesser extent graph structure learning. It deserves a serious referee and, with an ablation of the transfer channel and significance tests, could be a solid TMLR paper. If I were the editor, I would send it out. Recommendation: engage with it; require the ablation and error bars before acceptance.","headline":"Fresh problem setting, clean method, decent results—but the claimed structure-transfer channel is not isolated, and the experiments lack significance tests.","tokens_in":18657,"tokens_out":4873,"would_cite":true,"duration_ms":43354,"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":"FedGLS transfers graph-structure knowledge from clients with edges to graphless clients through a distilled feature encoder, letting feature-only clients learn usable local graphs and approach full-graph accuracy.","keywords":["federated learning","graph neural networks","graphless clients","graph structure learning","knowledge distillation","contrastive learning","node classification"],"falsifier":"Remove the knowledge-distillation loss from the feature encoder (or randomize the encoder before aggregation) and measure graphless-client accuracy: if it does not fall toward the feature-only baseline, the structure-transfer mechanism is not what drives FedGLS's gain. A complementary check is to hold out the true edge sets of graphless clients and test whether the adjacency matrices learned by FedGLS preserve the label-predictive edges.","tokens_in":17667,"feed_emoji":"🔗","tokens_out":7945,"duration_ms":70274,"temperature":0.7,"pith_summary":"Federated graph learning usually assumes every client has both node features and edge structure. FedGLS removes that assumption: it introduces 'graphless clients' that own only node features and shows they can still train effective GNNs by learning local graph structure from structure knowledge transferred by the other clients. The transfer works by giving every client a feature encoder (an MLP) that imitates its GNN's node embeddings through knowledge distillation, then federated-averaging both the GNN and the encoder parameters; the averaged encoder carries structure knowledge to graphless clients, where a local graph learner generates an adjacency matrix and is trained by a contrastive loss that aligns the GNN and encoder embeddings. Across five benchmark datasets, FedGLS consistently beats five baselines, including kNN-based structure construction, and approaches the accuracy of a system that uses the real graph structures.","feed_headline":"Federated learning works even when some clients have no graph edges","feed_subtitle":"A distilled feature encoder transfers structure from linked clients to feature-only ones.","key_machinery":"The load-bearing object is the feature encoder $h(\\cdot; \\phi)$, a feature-only MLP that is trained to match the GNN's node embeddings through the knowledge-distillation loss $\\mathcal{L}_{KD}^{(k)}$ of Eq. (11). Together with the GNN parameters $\\theta$, the encoder parameters $\\phi$ are aggregated on the server by the FedAvg-style update of Eq. (12), so the globally averaged encoder acts as the transport vehicle for structure knowledge. On each graphless client this averaged encoder supervises a graph learner $g(\\cdot; \\omega)$ that generates an adjacency matrix $S^{(k)}$ via an attentive encoder followed by a kNN-based adjacency processor; the graph learner is trained by the NT-Xent contrastive loss $\\mathcal{L}_{CL}^{(k)}$ of Eq. (9), which pulls together the GNN embeddings $Z^{(k)}$ and encoder embeddings $H^{(k)}$ for the same node and pushes apart embeddings of different nodes.","core_discovery":"The central claim is that graph-structure knowledge in a federated system can be moved from clients that have edges to clients that have none, using a feature encoder as a carrier. On each client the GNN produces node embeddings over the local graph, and an MLP feature encoder is distilled to reproduce those embeddings from features alone. The server averages both the GNN parameters and the encoder parameters with FedAvg, so the aggregated encoder becomes a global summary of structure-aware behavior. On a graphless client, a graph learner generates a sparse symmetric adjacency matrix from node features and is trained with a contrastive loss that maximizes agreement between the GNN embeddings computed on that generated graph and the encoder embeddings; this gives the graph learner an external structure signal to imitate. The result is that graphless clients learn usable adjacency matrices without ever seeing an edge, and the joint model's node-classification accuracy approaches that of training with true graph structure.","pith_inferences":["The same distillation-aggregation recipe could apply to other missing-data settings, for example clients that lack node features but have edges, by swapping which modality is distilled.","The contrastive objective that pulls GNN embeddings toward the encoder also discourages over-smoothing, so the accuracy gain on graphless clients may come partly from more stable GNN training rather than from faithful structure recovery; the paper's fixed-k sparsification suggests the learned structures are approximate.","A direct testable extension would be to use FedGLS's learned adjacency matrices as a warm start for structure refinement on graphless clients, or to measure how the fidelity of the transferred structure varies with the fraction of graphless clients.","The paper's fixed-k neighbor sparsification bounds the expressiveness of learned graphs; allowing the sparsity pattern to vary per node or be learned could improve recovery of underlying structure."],"forward_implications":["Graphless clients can train GNNs without ever observing edges, as long as other clients share structure knowledge through the federated aggregation of encoder and GNN parameters.","FedGLS outperforms kNN-based structure construction (Fed-GNNk) and feature-only baselines on all five datasets, with accuracy approaching that of Fed-GNN, which uses the true graph structures.","FedGLS converges faster than the best baseline Fed-GNNk on Cora and Flickr.","The framework's performance is stable across local epoch choices and graphless-client ratios, degrading gracefully as the fraction of graphless clients grows.","FedGLS adds only a small computational overhead on clients because the feature encoder is an MLP whose cost is much smaller than the GCN's."],"supporting_citations":[{"why":"Supplies the FedAvg aggregation rule (Eqs. 4 and 12) that merges client GNN and encoder parameters.","marker":"(McMahan et al., 2017)"},{"why":"Provides the knowledge-distillation mechanism used to train the feature encoder to imitate the GNN's embeddings.","marker":"(Hinton et al., 2015)"},{"why":"Defines the NT-Xent contrastive loss used by the graph learner on graphless clients.","marker":"(Chen et al., 2020a)"},{"why":"Shows that an MLP can be distilled from a GNN, the foundation for the feature-encoder design.","marker":"(Zhang et al., 2022)"},{"why":"Supplies the adjacency-processor operations (sparsification, symmetrization, normalization) used by the graph learner.","marker":"(Liu et al., 2022)"},{"why":"Provides evidence that feature-only and kNN-based structures underperform learned graph structures, motivating the transfer approach.","marker":"(Franceschi et al., 2019)"}],"fun_headline_variants":["Federated learning works for clients with no graph edges","Distilled encoder lets feature-only clients join federated GNNs","Graphless clients learn graph structure from others in federated setup","Federated GNNs get structure from feature-only clients via distillation","Sharing graph knowledge: federated learning with graphless clients"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework assumes that averaging the feature-encoder parameters across clients transfers usable graph-structure knowledge, and that a feature encoder trained by knowledge distillation preserves enough of that structure for a graphless client's graph learner to learn from; if the structure signal is lost in distillation or averaging, the contrastive objective has no useful teacher and FedGLS degenerates into feature-only training.","fun_headline_variants_meta":{"raw":{"variants":["Federated learning works for clients with no graph edges","Distilled encoder lets feature-only clients join federated GNNs","Graphless clients learn graph structure from others in federated setup","Federated GNNs get structure from feature-only clients via distillation","Sharing graph knowledge: federated learning with graphless clients"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000522,"raw_usage":{"total_tokens":2521,"prompt_tokens":935,"completion_tokens":1586,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":1498}},"tokens_in":551,"tokens_out":1586,"duration_ms":12087,"temperature":1.0,"reasoning_tokens":1498,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:39:07.289696+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Remove the knowledge-distillation loss from the feature encoder (or randomize the encoder before aggregation) and measure graphless-client accuracy: if it does not fall toward the feature-only baseline, the structure-transfer mechanism is not what drives FedGLS's gain. A complementary check is to hold out the true edge sets of graphless clients and test whether the adjacency matrices learned by FedGLS preserve the label-predictive edges.","supporting_citations":[{"cited_title":"Communication-efficient learning of deep networks from decentralized data","cited_arxiv_id":null,"evidence_quote":"Supplies the FedAvg aggregation rule (Eqs. 4 and 12) that merges client GNN and encoder parameters."},{"cited_title":"Learning discrete structures for graph neural networks","cited_arxiv_id":null,"evidence_quote":"Provides evidence that feature-only and kNN-based structures underperform learned graph structures, motivating the transfer approach."}],"review_version":1}