{"id":"b444b4f2-b0de-42dd-aef0-b2d5db9b8071","arxiv_id":"2501.10010","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"STAA uses wavelet-based node activity scores to bias a time-aware random walk, producing an augmented adjacency matrix that improves dynamic GNN performance.","lead":"This paper introduces STAA, a method that uses graph wavelet coefficients to identify nodes with noisy edges in dynamic graphs and adjusts a random walk so that recent noisy edges get less weight. It reports consistent gains over existing dynamic graph augmentation methods on node classification and link prediction.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section III-B's random walk probability 1−α−β_{t,u} can be negative under the stated hyperparameter ranges because β_{t,u}=δ·σ(τ̂) with δ∈(0,2] and α∈(0,1) is not clamped in Eq. (6).","rationale":"The reader's weakest assumption targets Observation (a), the wavelet-to-noise heuristic. That is a real validity concern for the paper's explanatory narrative, but it does not by itself undermine the algorithm's formal operation. The more load-bearing issue is that the random walk in Eq. (6) is not guaranteed to be a valid stochastic process under the paper's own hyperparameter ranges: β_{t,u} can exceed 1−α, making transition probabilities negative and potentially producing a diffusion matrix with negative entries. This is an internal inconsistency in the central construction, not merely an external heuristic. If the implementation silently clamps β, then the paper's equations misdescribe the method; if it does not clamp, the method as written is ill-defined. Both possibilities threaten reproducibility and the empirical central claim. The concrete test—checking the code or simulating with stated ranges—can settle which case holds. I therefore agree with the reader's overall CONDITIONAL verdict but identify a different, more formal concern as the primary reason to withhold unconditional acceptance.","tokens_in":9231,"tokens_out":7873,"duration_ms":80462,"concrete_test":"Inspect the released code or simulate Eq. (5) on a real dynamic graph with α=0.15, δ=2, and check max_{t,j} β_{t,j} across snapshots. If any β_{t,j} > 1−α, determine whether the implementation clamps β before forming L_t. If no clamp exists, compute X_t from Eq. (7) and test for negative entries; negative entries would confirm that the published equations do not define a valid probability diffusion. If a clamp exists, re-derive the update with min(β_{t,j}, 1−α) and verify that the reported AUC/Macro-F1 gains persist with the corrected equations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central algorithm is defined by the recursive update in Eq. (6), where the walker moves to a neighbor with probability 1−α−β_{t,u}. The text explicitly says β_{t,u} has an upper limit of 1−α, but no clamping is shown in Eq. (5), Eq. (6), or Eq. (7). Since β_{t,u}=δ·σ(τ̂_{t,j}) with δ up to 2, and σ(·) is the sigmoid, any node with τ̂ near zero gives β≈δ/2; for δ=2 this already exceeds 1−α for typical α<1. For nodes with τ̂>0, β approaches δ, making 1−α−β negative. Consequently, the diagonal matrix I−αI−β_{t,∧} in Eq. (6) can have negative entries, the described random walk is not a valid probability transition system, and L_t = I−Ã_t^T(I−αI−β_{t,∧}) may fail to be a nonsingular M-matrix. The resulting X_t may contain negative entries, so the 'diffusion matrix' is not a valid weighted adjacency matrix. The reader's concern about Observation (a) is legitimate but secondary: even if the wavelet-to-noise mapping is accepted, the walk as written is ill-defined for active nodes. This is an internal inconsistency: either the implementation clamps β (which the paper omits), or the equations do not match the described method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes STAA, a model-agnostic augmentation method for discrete-time dynamic graphs. It computes per-node spatiotemporal activity coefficients using graph wavelet coefficients and their rate of change, then biases a random walk with restart and time travel to produce a diffusion matrix that is used as an augmented adjacency matrix for dynamic GNN training. Experiments on seven datasets compare STAA with static and dynamic augmentation baselines on link prediction and node classification, reporting consistent improvements over no augmentation and over the dynamic methods TIARA and TGAC. The central claim is that suppressing recently preferred edges for active nodes reduces the influence of noise and improves dynamic graph learning.","tokens_in":9584,"tokens_out":5384,"duration_ms":53277,"significance":"If the algorithm is corrected and validated, the paper makes a useful contribution: it introduces a novel, principled mechanism for incorporating spatiotemporal activity into dynamic graph augmentation, with code and a reproducible experimental protocol (five runs, error bars, standard snapshot partitions). The use of graph wavelet coefficient change rates to detect temporally active nodes is a distinctive idea that goes beyond the temporal-locality assumption of prior methods. The experiments cover relevant baselines and datasets. However, the current formulation contains a load-bearing technical inconsistency in the random-walk transition probabilities, and the core Observation linking wavelet coefficients to noise susceptibility is not empirically established. These issues must be addressed before the claims can be accepted.","major_comments":[{"comment":"The transition probabilities are not well-defined as written. In Eq. (5), β_{t,u}=δ·σ(τ̂_{t,u}) with δ∈(0,2] and σ(·)∈(0,1), so β_{t,u} can exceed 1−α for many nodes. The text states that β has an upper limit of 1−α, but no clamping or reparameterization appears in Eqs. (5)-(7). Consequently, the 'random walk' probability 1−α−β_{t,u} can be negative, and L_t = I−Ã_t^T(I−αI−β_{t,∧}) need not be a nonsingular M-matrix. As a result, X_t can contain negative entries and is not guaranteed to be a valid weighted adjacency matrix. Please either clamp β to [0,1−α], reparameterize β as δ·σ(τ̂) with δ∈(0,1−α), or otherwise ensure 0≤β≤1−α, and then verify that the reported results are reproduced with the actual implementation.","section":"Section III-B, Eqs. (5)-(6)"},{"comment":"The method's entire activity coefficient rests on the unproven mapping from high-frequency wavelet coefficients to noise susceptibility. The manuscript offers intuition via the small example in Figure 2 and cites [18], but it provides no formal argument or direct empirical measurement that nodes with high high-frequency coefficients are actually more likely to be incident to noisy edges in real dynamic graphs. Because this Observation is the basis for the spatial activity term in Eq. (5), please provide empirical support—for example, inject synthetic noise into snapshots and show that high-frequency coefficients concentrate on corrupted nodes, or compare STAA's activity coefficients with ground-truth noisy-edge labels on a dataset with known noise.","section":"Section II, Observation (a)"},{"comment":"Even if β is clamped, the recursive update in Eq. (6) is not a standard stationary-distribution equation because the time-travel term β_{t,∧}x_{t−1,s} couples to the previous snapshot. The paper claims that x_{t,s} is a probability distribution, but with this extra term x_{t,s} is not obviously nonnegative or normalized, and no convergence proof for the recurrence is given. Since X_t is used as an augmented adjacency matrix, entrywise nonnegativity is essential. Please provide a proof or state explicit sufficient conditions (e.g., spectral radius of L_t^{-1}β_{t,∧} < 1, nonnegativity of L_t^{-1}) under which X_t is entrywise nonnegative and bounded.","section":"Section III-B, Eqs. (6)-(7)"}],"minor_comments":[{"comment":"The notation β_{t,∧} is confusing; using diag(β_{t,1},...,β_{t,n}) or B_t would be clearer and more standard.","section":"Section III-B, Eq. (6)"},{"comment":"No statistical significance tests are reported; with five runs, pairwise differences between some comparable methods (e.g., STAA vs. TIARA on several cells) may not be significant. Please add a significance test or at least error-bar analysis.","section":"Tables I-II"},{"comment":"The very large improvements on some datasets (e.g., Table I, WikiElec with EvolveGCN, 59.3→85.1) are striking and would benefit from a brief explanation or an ablation to rule out an artifact of the sparsification threshold or hyperparameter search.","section":"Section IV"},{"comment":"DropEdge and GDC results are taken from [11] rather than recomputed in the same environment; while this is acknowledged, a short note on the compatibility of the experimental setups would strengthen reproducibility.","section":"Section IV-A"},{"comment":"The hyperparameter ranges in the implementation details (δ∈(0,2]) directly conflict with the text's assertion that β has an upper limit of 1−α; this contradiction should be resolved in the text, not only in the equations.","section":"Section III-A, Eq. (5)"}],"recommendation":"major_revision","confidential_remarks":"The technical flaw in Eq. (6) is load-bearing but likely fixable; after clamping or reparameterizing β, the authors should rerun the experiments and report whether the improvements persist. The lack of any ablation isolating the spatial and temporal components also makes it difficult to attribute the gains to the paper's novelty; a revision should include such an ablation. The paper is within scope for the journal, and the code release is a positive factor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper's augmentation rule is genuinely new—node-specific activity coefficients computed from graph wavelet coefficients steer a TIARA-style random walk—and the empirical story is consistent across seven datasets and two tasks. But the method as written has an internal inconsistency that must be fixed before the theoretical framing holds up.\n\nThe problem is in Eq. (6). The walker moves with probability 1−α−β_{t,u}, and β_{t,u} is defined as δ·σ(τ̂) with δ up to 2 and σ the sigmoid. Nothing clamps β to 1−α, which the text claims as an upper limit. For any node with τ̂ near zero and δ=2, β≈1, so 1−α−β is negative for typical α<1; for positive τ̂ it gets worse. That means the diagonal matrix I−αI−β can have negative entries, L_t may not be a nonsingular M-matrix, and the resulting X_t can contain negative values. Either the implementation clamps β somewhere, in which case the paper needs to say so, or the equations do not describe a valid random walk. This is load-bearing, not cosmetic.\n\nWhat the paper does well: the spatiotemporal activity coefficient is a reasonable design, the random walk extension is a natural way to make time travel node-dependent, and the experiments are run five times with error bars. STAA beats NONE, TIARA, and TGAC on most metrics, and the code is promised. That is real evidence the idea has value.\n\nThe other soft spot is Observation (a): the claim that high-frequency wavelet coefficients indicate noise-prone critical nodes. The paper gives a small illustrative example but no formal argument or direct empirical check on real dynamic graphs. That is a genuine limitation, though for an applied augmentation paper it is secondary if the empirical results hold.\n\nMinor issues: TGAC was reproduced by the authors and DropEdge/GDC results are borrowed from the TIARA paper, so the baseline comparison is not fully apples-to-apples. Hyperparameter tuning is otherwise standard.\n\nNet: the empirical claim is plausible and the idea is worth engaging with, but the paper in its current form has a gap in the math that a referee should insist be fixed. I would send it to peer review, with the request that the authors either clamp β explicitly or restrict δ, and show that X_t entries are non-negative in practice.","headline":"STAA is a clever wavelet-guided extension of TIARA with consistent empirical wins, but the random walk probabilities as written can go negative and need a clamp before the method is fully coherent.","tokens_in":10096,"tokens_out":2726,"would_cite":false,"duration_ms":28885,"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":"A wavelet-based activity score identifies noise-prone nodes, and a biased random walk turns that score into a diffusion matrix that improves dynamic graph learning.","keywords":["dynamic graph learning","graph augmentation","graph wavelet transform","random walk diffusion","graph neural networks","noise robustness","temporal locality","link prediction"],"falsifier":"Build a dynamic graph with injected noise placed deliberately on nodes with smooth, stable neighborhoods (low high-frequency coefficients) and put stable patterns on irregular nodes; then STAA's diffusion matrix should assign high weight to noise if the activity signal is working as claimed, and link prediction should degrade relative to no augmentation. Alternatively, inspect the learned diffusion matrix on a real dataset and check directly whether noisy edges receive lower weight than stable edges for the same node.","tokens_in":9055,"feed_emoji":"🕸️","tokens_out":5822,"duration_ms":52914,"temperature":0.7,"pith_summary":"Dynamic graphs contain noisy edges that appear and vanish for reasons unrelated to the underlying process, and most augmentation methods assume recent edges are more important—an assumption that lets such noise dominate learning. This paper proposes STAA, a model-agnostic augmentation that scores every node's spatiotemporal activity from graph wavelet coefficients and their change rate over snapshots, then uses a random walk with node-specific time-travel probabilities to build a diffusion matrix that down-weights edges attached to active nodes. The diffusion matrix replaces the adjacency matrix when training any discrete-time dynamic graph neural network. Experiments on seven datasets report that STAA beats no augmentation, static augmentations, and the temporal-locality baselines TIARA and TGAC on both link prediction and node classification.","feed_headline":"Activity-aware walks suppress noisy edges in dynamic graphs","feed_subtitle":"Wavelet-based activity steers random walks toward stable edges, lifting GNN accuracy across seven datasets.","key_machinery":"The central object is the SpatioTemporal Activity-Aware Random Walk Diffusion matrix $X_t$, generated by the STAARW random walk. On each snapshot $G_t$, a walker moves to a neighbor with probability $1-\\alpha-\\beta$, restarts at the seed with probability $\\alpha$, or jumps to the same node in the next snapshot with probability $\\beta$, where $\\beta$ is the node's spatiotemporal activity coefficient. The coefficient combines the normalized high-frequency wavelet coefficient (spatial activity) and the normalized rate of change of the low-frequency coefficient over a window of $W$ snapshots (temporal activity). The stationary distribution of this walk, stacked over all seed nodes, yields the recurrence $X_t = \\alpha L_t^{-1} + L_t^{-1} \\beta_t X_{t-1}$; the resulting matrix is sparsified by a threshold $\\rho$ and then used as the augmented adjacency matrix. The machinery converts a continuous spectral estimate of noise-proneness into concrete edge weights before any GNN forward pass.","core_discovery":"The paper's central claim is that noise in dynamic graphs can be identified, without labels, from node-level spectral behavior: nodes whose neighborhoods are structurally irregular (high high-frequency graph wavelet coefficients) and whose low-frequency coefficients change sharply across snapshots are the nodes most likely to carry spurious recent edges. STAA condenses these two signals into a spatiotemporal activity coefficient for each node and time step, and uses it to bias a random walk on the snapshot sequence. The walk either moves to a neighbor, restarts at the seed, or time-travels to the next snapshot with a probability proportional to the node's activity; solving this walk gives a diffusion matrix whose entries re-weight the graph so that active nodes' recent-edge preference is suppressed. This matrix is then used in place of the original adjacency matrix for GNN training. The paper argues this is why STAA outperforms augmentations built purely on temporal decay or spatial denoising.","pith_inferences":["I would expect STAA's advantage to shrink on graphs where node activity is nearly uniform, because then the temporal-locality bias and the activity bias coincide and the extra machinery mainly adds variance.","A direct test of the noise-identification claim would be to compare STAA's diffusion weights against an oracle that knows which edges are injected noise, isolating how much of the gain comes from correct identification versus from generic smoothing.","The same wavelet-based activity signal could be reused outside augmentation, for instance as an edge-weighting scheme for static GNN training or as a node-selection heuristic for active learning, though the paper does not explore these uses."],"forward_implications":["Dynamic GNNs using STAA's diffusion matrix should become less sensitive to spurious single-snapshot edges, since the random walk shifts weight from an active node's newest edges to its earlier, more stable edges.","The augmentation is plug-and-play: the paper demonstrates it with GCN, GCRN, and EvolveGCN, so any discrete-time dynamic GNN can consume the diffusion matrix without architectural changes.","STAA requires no labels, noise annotations, or generative models; the activity coefficient is computed purely from adjacency matrices and node degrees across snapshots.","If the reported gains generalize, augmentation methods built solely on temporal locality will need a node-activity component rather than a uniform recent-edge emphasis."],"supporting_citations":[{"why":"Supplies the TIARA time-aware random walk diffusion baseline that STAA extends and compares against.","marker":"[11]"},{"why":"Supplies the TGAC temporal-locality augmentation baseline used in the comparison.","marker":"[12]"},{"why":"Defines the graph wavelet transform that STAA uses to compute node activity.","marker":"[16]"},{"why":"Provides the wavelet filter kernel and the wavelet-based dynamic network analysis that the coefficient computation follows.","marker":"[17]"},{"why":"Supports the claim that nodes with high-frequency graph signal content are susceptible to noise.","marker":"[18]"},{"why":"Random walk with restart is the base process that STAARW generalizes.","marker":"[19]"},{"why":"Gives the cross-snapshot random walk diffusion formulation that STAARW adapts with activity-based biasing.","marker":"[20]"},{"why":"Supplies the node classification datasets and evaluation protocol used in experiments.","marker":"[24]"}],"fun_headline_variants":["Wavelet activity steers random walks to suppress noisy edges","STAA: Spectral activity flags noisy edges without labels","Spatiotemporal activity reweights graph to boost GNN accuracy","Noisy edge detection via wavelet-driven random walk diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that high high-frequency wavelet coefficients really do mark nodes whose edges are likely to be noise, and that low-frequency coefficient changes really do mark temporally unstable nodes; if that spectral-to-noise mapping fails on a given graph, STAA will down-weight the wrong edges.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet activity steers random walks to suppress noisy edges","STAA: Spectral activity flags noisy edges without labels","Spatiotemporal activity reweights graph to boost GNN accuracy","Noisy edge detection via wavelet-driven random walk diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000242,"raw_usage":{"total_tokens":1491,"prompt_tokens":876,"completion_tokens":615,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":560}},"tokens_in":492,"tokens_out":615,"duration_ms":6834,"temperature":1.0,"reasoning_tokens":560,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:23:20.231312+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a dynamic graph with injected noise placed deliberately on nodes with smooth, stable neighborhoods (low high-frequency coefficients) and put stable patterns on irregular nodes; then STAA's diffusion matrix should assign high weight to noise if the activity signal is working as claimed, and link prediction should degrade relative to no augmentation. Alternatively, inspect the learned diffusion matrix on a real dataset and check directly whether noisy edges receive lower weight than stable edges for the same node.","supporting_citations":[{"cited_title":"Time-aware random walk diffusion to improve dynamic graph learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the TIARA time-aware random walk diffusion baseline that STAA extends and compares against."},{"cited_title":"Temporal graph representation learning with adaptive augmentation contrastive,","cited_arxiv_id":null,"evidence_quote":"Supplies the TGAC temporal-locality augmentation baseline used in the comparison."},{"cited_title":"Wavelet-based visual analysis of dynamic networks,","cited_arxiv_id":null,"evidence_quote":"Provides the wavelet filter kernel and the wavelet-based dynamic network analysis that the coefficient computation follows."},{"cited_title":"Discrete signal processing on graphs: Frequency analysis,","cited_arxiv_id":null,"evidence_quote":"Supports the claim that nodes with high-frequency graph signal content are susceptible to noise."},{"cited_title":"Localization in Seeded PageRank","cited_arxiv_id":"1509.00016","evidence_quote":"Random walk with restart is the base process that STAARW generalizes."},{"cited_title":"Time-aware Random Walk Diffusion to Improve Dynamic Graph Learning","cited_arxiv_id":"2211.01214","evidence_quote":"Gives the cross-snapshot random walk diffusion formulation that STAARW adapts with activity-based biasing."},{"cited_title":"Spatio-temporal attentive rnn for node classification in temporal attributed graphs,","cited_arxiv_id":null,"evidence_quote":"Supplies the node classification datasets and evaluation protocol used in experiments."}],"review_version":1}