{"id":"e0fa66f3-aa85-4755-a529-d83f64af478f","arxiv_id":"2506.13099","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"DyGC, the first framework for dynamic graph condensation, synthesizes a small temporal graph that preserves enough spatiotemporal structure to train dynamic GNNs with up to 1846 times speedup and around 96 percent fidelity.","lead":"This paper introduces DyGC, a method that compresses a large dynamic graph into a tiny synthetic graph while keeping most of its predictive power for graph neural networks. If it holds up, it could make training on evolving networks such as social or transaction graphs much faster and cheaper.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Matching MMD in the hand-crafted linear field (Eq. 6) is assumed to transfer to arbitrary DGNN accuracy; no argument shows this field preserves the information each DGNN needs. If a DGNN exploits structure Eq. 6 discards, the condensed graph underperforms despite low MMD.","rationale":"I read the paper as making an empirical claim: DyGC produces a small dynamic graph that, when used to train a DGNN, yields accuracy close to training on the full graph. The mechanism is spatiotemporal distribution matching in the state field of Eq. 6. The reader's verdict (CONDITIONAL) matches my reading: the method is plausible and well-evaluated, but the central mechanism rests on an unproven transfer property of a hand-chosen linear field. This is the single most load-bearing concern because if Eq. 6 discards information that a downstream DGNN needs, the entire condensation objective is misaligned with the task. The existing experiments on five backbones (T-GCN, GCRN, STGCN, DySAT, ROLAND) provide real support—fidelity of 86.4–98.7% is not trivial—and the ablation shows DGM contributes 2.0–8.5%, so the field is doing work. However, all backbones are snapshot-based and operate in roughly the same linear-diffusion-plus-RNN/attention regime; none requires, say, edge-level features, temporal velocity, or long-range nonlinear interactions. I do not see evidence that the field is architecture-independent. I also noticed that Proposition 3.1 is stated with an 'iff' that is false (matched dependent chains can have equal joint and marginal KL divergences), but I regard this as a motivational/theoretical flaw, not the main load-bearing risk; the method's validity does not depend on that proposition. For these reasons I agree with the reader's weakest assumption and keep the verdict CONDITIONAL (no change). A constructive fix would be to replace the absolute claim 'faithfully preserving inherent spatiotemporal characteristics' with a scope-limited claim, and to add the transfer stress test above.","tokens_in":22301,"tokens_out":8664,"duration_ms":100029,"concrete_test":"Design a transfer stress test: with logit alignment disabled (γ=0), condense DBLP/Reddit using the pure Eq. 6/7 MMD objective, then evaluate on T-GCN, DySAT, and a nonlinear temporal DGNN (per-snapshot GAT + Transformer) not in the current set. Also run the same procedure with the state field replaced by a nonlinear one (GAT layer + GRU). If the linear-field graph loses more than 2% fidelity on the nonlinear backbone relative to the nonlinear-field graph, or if its fidelity drops sharply while the Eq. 6 MMD stays low, then the Eq. 6 field does not transfer to architectures outside its linear-diffusion span.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DyGC's distribution matching loss (Eq. 7) compares real and synthetic graphs only through the state evolving field H defined in Eq. 6: H(t,k) = αH(t-1,k) + (1-α)M(t)H(t,k-1), with boundary states H(1,k) = (M(1))^{k-1}X(1) and H(t,1) = (1-α)X(t) + αX(t-1). This is a fixed linear spatiotemporal diffusion: α is fixed at 0.5, M(t) is the normalized adjacency (with self-loops), and K is small (K ≤ 3 in the sensitivity study). The condensed graph is optimized so that class-wise MMD between H_T and H_S is small, plus a logit-alignment term. The load-bearing step is the implicit claim that small MMD in this particular hand-chosen field implies that any DGNN—e.g., DySAT's self-attention, ROLAND's hierarchical states, or an attention-plus-Transformer DGNN—will have similar predictive behavior on the real and condensed graphs. Eq. 6 is a linear map of the graph sequence; it cannot represent nonlinear feature interactions, learned attention weights, or higher-order temporal statistics. No bound of the form 'DGNN loss difference ≤ f(MMD)' is derived; the transfer is purely empirical. The experiments show strong fidelity on five backbones, which supports the claim for those backbones, but the stated contribution—'faithfully preserving the inherent spatiotemporal characteristics'—is stronger than what is established. A related rigor issue is Proposition 3.1: as stated ('iff' both sequences are independent), it is false; equality of joint and marginal KL divergences can hold for dependent sequences with matched conditionals. This does not by itself sink the method, but it shows the theoretical framing around temporal dependence is not reliable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DyGC, a method for dynamic graph condensation that synthesizes a small discrete-time dynamic graph from a large one. It uses a spiking-neuron-based structure generator (SSG) and a distribution matching objective based on maximum mean discrepancy (MMD) in a recursively constructed spatiotemporal 'state evolving field' (Eq. 6), plus a logit alignment term from a pretrained DGNN. Experiments on four dynamic graph datasets and five DGNN backbones report that DyGC outperforms per-snapshot static condensation and coreset baselines, retaining up to 96.2% fidelity at 0.5% of the original graph size and achieving large training speedups.","tokens_in":22742,"tokens_out":7230,"duration_ms":81677,"significance":"If the empirical results hold, DyGC is a practically valuable contribution to data-centric dynamic graph learning. The paper provides broad evidence: four datasets of varying scale, five diverse DGNN backbones, ablations of each module, temporal continuity analysis, and explicit storage and training-speedup quantification. The setup is described in sufficient detail in Appendix E, and Algorithm 1 in Appendix D clarifies the pipeline. However, the theoretical motivation contains a false proposition (Prop 3.1), the headline fidelity claim is based on the best metric and ratio (Micro-F1 on Arxiv at the largest ratio) while Macro-F1 fidelity drops to 86.8% in the same cell, and the transfer from the hand-designed linear state field to arbitrary DGNN accuracy is an unproven modeling assumption rather than a derived guarantee.","major_comments":[{"comment":"The claimed 'iff' in Proposition 3.1 is false. If the two dynamic graphs have identical joint distributions (e.g., P = Q with perfectly dependent snapshots), then the joint KL divergence and the sum of marginal KL divergences are both zero, so equality holds even though neither sequence is temporally independent. The proof in Appendix C only constructs one configuration where non-independence leads to inequality; it does not establish necessity. The motivational point that marginal-only alignment is insufficient can be preserved, but the formal statement and proof must be corrected, for example by restating the result as an impossibility example rather than an equivalence.","section":"Section 3, Prop. 3.1 and Appendix C"},{"comment":"The distribution matching loss compares only the hand-designed linear state evolving field H, defined as a fixed spatiotemporal diffusion, and no argument is given that small MMD in this field transfers to small error for arbitrary DGNNs (e.g., DySAT's self-attention or ROLAND's hierarchical states). The paper's claim that DyGC 'faithfully preserves the inherent spatiotemporal characteristics' is therefore stronger than what is established; generalization across the five backbones is empirical and may depend on the chosen field. In addition, the logit alignment in Section 4.3 uses T-GCN as the only pretrained architecture; the paper should clarify whether the same condensed graph is used for all backbones and discuss the effect of this task-model mismatch.","section":"Section 4.2, Eq. (6) and Eq. (7)"},{"comment":"The headline claim 'retains up to 96.2% DGNN performance' is the Micro-F1 fidelity for T-GCN on Arxiv at the largest condensation ratio, while the Macro-F1 fidelity for the same cell is only 86.8%. The 'up to' phrasing is technically correct but obscures that Macro-F1 fidelity is substantially lower. Please report both metrics in the abstract/conclusion or explicitly discuss the fidelity range (86.4%–98.7%) and the Micro/Macro gap, which is relevant for class-imbalanced applications.","section":"Abstract and Table 2"}],"minor_comments":[{"comment":"The notation D_KL(P_{G_{1:T}} \\| Q_{\\tilde{G}_{1:T}}) in the proposition is not formally defined for distributions over different graph sequences; clarify that Q denotes the distribution of the synthetic dynamic graph.","section":"Section 2"},{"comment":"The update equation for \\hat{U}^{(t)}_{ij} uses U^{(t-1)}_{ij} on the right-hand side but does not explicitly define the integration time step or the reset mechanism; please provide the full discretized LIF update.","section":"Section 4.1, Eq. (4)"},{"comment":"The kernel function \\psi over tensors of shape K × T × d may suffer from high-dimensional MMD; the paper should specify how the RBF kernel bandwidth is chosen and whether features are normalized.","section":"Section 4.2, Eq. (8)"},{"comment":"The y-axis label 'Scale = ln(epoch/sec)' is inconsistent with the plotted values (which appear to be speedup factors); replace it with a clearly defined speedup metric.","section":"Figure 4"},{"comment":"The 'Whole' column is a full-graph upper bound, not a baseline method; it should be labeled accordingly to avoid confusion.","section":"Table 1 and 2"},{"comment":"The sensitivity analysis covers only the number of spatial propagation steps K; please add sensitivity studies for the temporal coefficient \\alpha, the membrane decay \\tau_m, the firing threshold U_th, and the logit-alignment weight \\gamma.","section":"Appendix E.4"}],"recommendation":"major_revision","confidential_remarks":"The empirical study is broad and the central method appears effective in the tested settings, but the false theoretical proposition and the overclaiming abstract need to be addressed before publication. No ethical concerns. The paper appears to be a preprint with a NeurIPS checklist; if it is a journal submission, the authors should adapt the formatting accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on arXiv:2506.13099. DyGC is genuinely the first condensation method for discrete-time dynamic graphs, and the empirical contribution is real. On four datasets (DBLP, Reddit, Arxiv, Tmall) and five DGNN backbones, it beats per-snapshot static condensation and coreset baselines by consistent margins at extreme compression ratios, and the storage/training speedups are large. The spiking structure generation mechanism is new and the ablations show it contributes. I believe the experimental core.\n\nThe soft spots are mostly in the framing. Proposition 3.1, the 'iff' claim about KL divergence, is false as stated. Equality of joint and marginal KL can hold with temporally dependent snapshots when the reference process has degenerate support and the divergence is measured only on that support; the proof's counterexample only demonstrates one direction. The temporal independence conclusion doesn't follow. This doesn't sink the method, but the paper uses the proposition to motivate the matching objective, and it should be corrected or removed.\n\nSecond, the state evolving field is a fixed linear spatiotemporal diffusion (Eq. 6). The paper provides no argument that matching MMD in this hand-crafted field transfers to arbitrary DGNN predictions. The five-backbone experiments are evidence for those backbones, including attention-based DySAT and hierarchical ROLAND, so the empirical transfer is better than nothing. But the abstract's 'faithfully preserving' claim is stronger than what is established. A short discussion of where the linear field could fail would be honest.\n\nThird, the headline fidelity number is the best case. On Arxiv at ratio 0.05%, Micro-F1 fidelity is 96.2% but Macro-F1 fidelity is 86.8% for the same setting. The paper should present a range or both metrics in the abstract and intro.\n\nMinor issues: code is promised but not actually visible in this submission; hyperparameters like tau_m, U_th, and beta are not fully specified; and the baselines are only static/coreset. Since this is the first dynamic method, that baseline set is defensible, though a simple temporal-random-sampling baseline would strengthen the comparison.\n\nIn sum: the central empirical claim holds up. This is a useful method paper for researchers working on graph condensation or dynamic graph learning, with a clear first-mover contribution. I'd send it to serious peer review. The reviewer asks should focus on the false Proposition 3.1, the transfer claim, and the fidelity reporting; these are fixable without new experiments.","headline":"First dynamic graph condensation method with broad, convincing experiments, but a false theoretical proposition and over-claimed fidelity need fixing.","tokens_in":23279,"tokens_out":11502,"would_cite":true,"duration_ms":119295,"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":"This paper introduces DyGC, the first framework for dynamic graph condensation, which distills a time-evolving graph into a compact synthetic graph that preserves spatiotemporal structure; a DGNN trained on 0.5% of the data keeps up to…","keywords":["dynamic graph condensation","spatiotemporal distribution matching","spiking neural networks","graph condensation","dynamic graph neural networks","state evolving field","temporal node classification","maximum mean discrepancy"],"falsifier":"Take the real and condensed graphs for a small dataset and estimate the KL divergence between their full joint distributions over snapshots; if the joint divergence remains large while per-snapshot marginal divergences are near zero, the state evolving field has failed to capture the temporal dependencies that DyGC claims to preserve.","tokens_in":22113,"feed_emoji":"⚡","tokens_out":7086,"duration_ms":73376,"temperature":0.7,"pith_summary":"Dynamic graphs record how connections and node attributes change over time, but they are bulky and expensive to train on. This paper introduces DyGC, which compresses an entire sequence of graph snapshots into a tiny synthetic dynamic graph that a dynamic graph neural network (DGNN) can train on instead. The central claim is that the compressed graph can preserve the real graph's spatiotemporal distribution closely enough that a DGNN trained on it keeps most of its accuracy: up to 96.2% of full-graph performance using only 0.5% of the original graph size, with training speedups up to 1,846 times. The guiding insight is that static graph condensation fails on dynamic graphs because matching each snapshot separately ignores temporal dependencies, so DyGC aligns the joint spatiotemporal state rather than per-snapshot marginals. A sympathetic reader would care because this makes large-scale time-evolving graph learning far more data-efficient.","feed_headline":"Condense a dynamic graph to 0.5% and keep 96% of its accuracy","feed_subtitle":"New framework distills time-evolving graphs so dynamic GNNs train up to 1,846x faster with near-full accuracy.","key_machinery":"The three load-bearing pieces are: (1) the spiking structure generator, a leaky integrate-and-fire mechanism (Eqs. 4–5) that turns node features into a temporally evolving binary adjacency matrix, with surrogate-gradient training to handle the non-differentiable firing threshold; (2) the state evolving field (Eq. 6), a recursive operator that mixes spatial message passing $\\mathbf{M}^{(t)}$ with temporal smoothing across snapshots, producing a tensor of node states that represents the graph's spatiotemporal dynamics; (3) class-wise maximum mean discrepancy on that field, which implicitly compares all moments of the state distributions and, together with logit alignment, drives the optimization of the synthetic graph. Together they convert the intractable problem of matching two dynamic-graph distributions into matching finite state tensors in a fixed semantic space.","core_discovery":"DyGC condenses a discrete-time dynamic graph $\\mathcal{T}=(\\mathbf{A},\\mathbf{X},\\mathbf{Y})$ into a small synthetic graph $\\mathcal{S}$ by jointly optimizing synthetic node features and an evolving binary adjacency structure so that the spatiotemporal distribution of $\\mathcal{S}$ matches that of $\\mathcal{T}$. Structure is produced by a spiking structure generation module, which treats inter-node affinity like the membrane voltage of a leaky integrate-and-fire neuron: affinity integrates past interaction memory with current feature-based stimulus, and an edge appears when the voltage crosses a threshold, with a soft reset after firing. Distribution matching is done in a 'state evolving field' $\\mathbf{H}^{(t,k)}=\\alpha\\mathbf{H}^{(t-1,k)}+(1-\\alpha)\\mathbf{M}^{(t)}\\mathbf{H}^{(t,k-1)}$, a recursive blend of spatial message passing and temporal smoothing, and the discrepancy between real and synthetic fields is measured class-wise with maximum mean discrepancy, plus a logit alignment term from a pretrained DGNN. On DBLP, Reddit, Arxiv, and Tmall, condensed graphs from DyGC outperform snapshot-wise static condensation baselines across five DGNN architectures, with reported performance fidelity of 86.4%–98.7% and large reductions in storage and training time.","pith_inferences":["The state evolving field is a linear blend of one-step spatial diffusion and one-step temporal smoothing; architectures that rely on long-range temporal memory (for example recurrent or state-space DGNNs) may carry information the field never represents, which would limit the claimed fidelity beyond the five tested backbones.","The discrete-time snapshot setting is a genuine scope restriction: event-based or continuous-time graphs require a different condensation objective, and the paper's own appendix acknowledges that discretization can lose temporal precision.","The method is evaluated on temporal node classification; for link prediction or anomaly detection, the information a condensed graph must preserve is different, so the same spatiotemporal matching objective would likely need re-design rather than simple reuse.","The spiking structure generator is a means to produce discrete evolving edges, not the only one; the transferable contribution is the joint spatiotemporal distribution-matching objective, which could be paired with other differentiable structure generators."],"forward_implications":["A DGNN trained on a DyGC-condensed graph can reach 86.4%–98.7% of its full-graph accuracy on temporal node classification across five architectures, so expensive training can run on drastically smaller data.","Snapshot-wise application of static condensation methods to dynamic graphs is provably insufficient: Proposition 3.1 shows that matching per-snapshot marginals only aligns the joint distribution when snapshots are temporally independent, which real dynamic graphs are not.","Condensation time is lower than static methods in most settings and grows more slowly with condensation ratio, because DyGC processes the whole snapshot sequence jointly rather than repeating condensation per snapshot.","The spiking structure generator yields discrete, unweighted edges whose consecutive-snapshot Jaccard similarity tracks that of real graphs, indicating structural continuity rather than rigidity."],"supporting_citations":[{"why":"Supplies the static graph condensation baseline that DyGC extends to the temporal setting.","marker":"[15]"},{"why":"Provides the logits-alignment idea and a strong static condensation baseline that DyGC generalizes to dynamic graphs.","marker":"[37]"},{"why":"Provides the maximum mean discrepancy metric used for the fine-grained spatiotemporal state alignment in Eq. 7.","marker":"[7]"},{"why":"Defines the leaky integrate-and-fire neuron model that underlies the spiking structure generation mechanism.","marker":"[12]"},{"why":"Supplies the surrogate-gradient technique that makes the discrete spike-based structure generator trainable.","marker":"[4]"},{"why":"T-GCN is the primary DGNN architecture used for the main comparisons and for generating soft labels in logit alignment.","marker":"[44]"},{"why":"Provides the four dynamic graph datasets (DBLP, Reddit, Arxiv, Tmall) used in the experiments.","marker":"[20]"},{"why":"A distribution-matching static condensation baseline that DyGC outperforms on dynamic graphs.","marker":"[21]"}],"fun_headline_variants":["Dynamic graphs condensed to 0.5% size with 96% accuracy","Squeeze dynamic graphs to 0.5% size, keep 96% accuracy","Dynamic graph condensation: 0.5% size, 96% accuracy","Condense dynamic graphs 200x, keep 96% accuracy","Dynamic graph condensation: 1846x faster training, 96% accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The condensed graph is optimized to match the real graph inside a hand-built 'state evolving field' that mixes one-step spatial and temporal message passing; if that field discards information a particular dynamic GNN needs, the condensed graph will underperform for that architecture.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic graphs condensed to 0.5% size with 96% accuracy","Squeeze dynamic graphs to 0.5% size, keep 96% accuracy","Dynamic graph condensation: 0.5% size, 96% accuracy","Condense dynamic graphs 200x, keep 96% accuracy","Dynamic graph condensation: 1846x faster training, 96% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001477,"raw_usage":{"total_tokens":5988,"prompt_tokens":1053,"completion_tokens":4935,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":669,"completion_tokens_details":{"reasoning_tokens":4843}},"tokens_in":669,"tokens_out":4935,"duration_ms":37451,"temperature":1.0,"reasoning_tokens":4843,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:36:44.422355+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the real and condensed graphs for a small dataset and estimate the KL divergence between their full joint distributions over snapshots; if the joint divergence remains large while per-snapshot marginal divergences are near zero, the state evolving field has failed to capture the temporal dependencies that DyGC claims to preserve.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the logits-alignment idea and a strong static condensation baseline that DyGC generalizes to dynamic graphs."},{"cited_title":"Incorporating Learnable Membrane Time Constant to Enhance Learning of Spiking Neural Networks","cited_arxiv_id":"2007.05785","evidence_quote":"Supplies the surrogate-gradient technique that makes the discrete spike-based structure generator trainable."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"T-GCN is the primary DGNN architecture used for the main comparisons and for generating soft labels in logit alignment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the four dynamic graph datasets (DBLP, Reddit, Arxiv, Tmall) used in the experiments."}],"review_version":1}