{"id":"630f5b7d-dd5a-4393-931a-284066ff6eb8","arxiv_id":"2506.22004","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"GKNet jointly learns a diffusion-style graph state model and a graph-filtered observation model, enabling accurate graph time-series tracking from scarce and partially observed data.","lead":"This paper introduces GKNet, a neural network that learns both how a signal spreads over a network and how sensors observe it, using a graph-based state-space model. It matters because it can forecast or fill in network sensor data such as traffic or water levels with very little training data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (7) is not the Euler discretization of the stated graph SPDE: the identity term is missing, and this error propagates into the Kalman prediction, EM objective, training loss, and water-network state equation, so the central claim that GKNet is a principled unrolling of the graph Kalman…","rationale":"The central claim of the paper is that GKNet is a principled deep learning architecture obtained by unrolling a graph Kalman filter whose state model is the Euler-discretized graph SPDE of Eq. (6). The reader's weakest-assumption analysis correctly identifies Eq. (7) as the point where this derivation breaks: the Euler step is missing the identity term, so the discrete state transition is not the discretization of the stated SPDE. This single error is load-bearing because it propagates into the Kalman prediction, the EM objective, the training loss, and the water-network modification, and it contradicts the paper's own Remark 1. I checked whether this could be dismissed as a harmless notational convention: it cannot, because the transition matrix -cL produces a different mean evolution and different covariance recursions than I-cL, and the likelihood and loss terms are written explicitly in terms of the erroneous transition. The empirical sections cannot rescue the derivation as written, since the experiments are presented as validating this particular state-space model and its unrolled architecture. I therefore agree with the reader's verdict: the paper's central mathematical support is invalid as stated, and the claim that GKNet is a principled unrolling of the graph Kalman filter is not established. The appropriate disposition is the same as the reader's rejection, so I do not propose changing the verdict. An honest path to acceptance would require correcting the discretization, updating all recursions and losses accordingly, and re-running the experiments before the central claim can be evaluated.","tokens_in":20966,"tokens_out":3239,"duration_ms":37483,"concrete_test":"Independently re-derive the Euler discretization of Eq. (6) and replace Eq. (7) with x_{t+1}=(I-cL)x_t+Bdiag(alpha)w_t. Then update the Kalman prediction (17) to x_t^-=(I-cL)x_{t-1} and P_t^-=(I-cL)P_{t-1}^+(I-cL)^T+Bdiag(alpha)^2B^T, and propagate the correction through Eqs. (16), (35), and (39). As an analytical check, simulate the mean recursion of the original and corrected discrete models against the SPDE solution mean exp(-cLt)mu0; the corrected recursion should match it and the original recursion should not. As an empirical check, rerun the METR-LA low-data forecasting and the linear tracking experiments with the corrected recursions and report nRMSE; if the reported low-data advantage disappears or requires materially different learned parameters, the original results depend on the missing identity term.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (7) states x_{t+1} = -cLx_t + Bdiag(alpha)w_t and calls this the first-order Euler discretization of d x_t = -cL x_t dt + Bdiag(alpha)d beta_t. Explicit Euler with a unit step gives x_{t+1} = x_t - cL x_t + Bdiag(alpha)w_t = (I-cL)x_t + Bdiag(alpha)w_t. The identity term x_t is missing. This is not a notational quibble: the discrete transition matrix is -cL instead of I-cL, so the deterministic part does not approximate the heat-diffusion semigroup exp(-cL). The same missing identity appears in the state-space model (12), in the Kalman prediction (17) via x_t^- = -cLx_{t-1} and P_t^- = c^2 L P_{t-1}^+ L^T + Q, in the negative log-likelihood (16) through the term (x_t + cLx_{t-1}), in the GKNet loss (35), and in the water-network state equation (39). This also contradicts Remark 1, whose mean evolution is exp(-cLt)mu0; the discrete model as written has mean evolution (-cL)^t mu0, which neither matches the continuous solution nor is stable in the same way. Because the paper's central claim is that GKNet is a principled model-based unrolling of a graph Kalman filter derived from this SPDE, the presented derivation does not support the claim. The empirical results may still indicate a useful graph recurrent architecture, but they are not evidence for the specific state-space derivation as written. The reader's weakest-assumption analysis identifies the correct load-bearing flaw.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a graph state-space model for time-varying graph signals, with a latent process driven by a graph SPDE with edge-localized noise and observations given by a masked graph filter. The authors first learn parameters and latent states via an EM algorithm based on Kalman smoothing, then extend this into GKNet, a model-based deep network that replaces the Kalman gain with a graph filter and estimates the model parameters with a recurrent inference module. The method is evaluated on traffic forecasting, weather interpolation, synthetic tracking, and a stormwater-network case study, with emphasis on low-data regimes.","tokens_in":21417,"tokens_out":12105,"duration_ms":119324,"significance":"If the theoretical derivation were correct, the paper would offer a useful model-based deep learning approach combining interpretability with scalability, and the extensive experiments, especially the water-network case study, would be valuable. The attempt to connect an SPDE-based graph kernel to recursive Kalman inference is a worthwhile direction, and the low-data empirical results are potentially interesting. However, the load-bearing derivations contain fundamental errors, so the significance is not realized in the current form: the empirical results cannot be attributed to the claimed SPDE state-space model without correcting and re-running the core derivation.","major_comments":[{"comment":"Equation (7) asserts that x_{t+1} = -cL x_t + Bdiag(alpha)w_t is the first-order Euler discretization of d x_t = -cL x_t dt + Bdiag(alpha)d beta_t. Explicit Euler with unit step gives x_{t+1} = x_t - cL x_t + ... = (I - cL)x_t + ..., so the identity term is missing. This is not a typo: the same recurrence appears in the state-space model (12), the negative log-likelihood (16), the Kalman prediction (17), the smoother gains (19)-(22), the GKNet loss (35), and the water-network state equation (39). It also contradicts Remark 1, where the continuous mean is exp(-cL t)mu0; the discrete model has mean (-cL)^t mu0, which neither approximates the heat semigroup nor is stable for general c. Since the paper's central claim is that GKNet is a principled unrolling of a graph Kalman filter derived from this SPDE, the derivation does not support the claim as written.","section":"II-A, Eq. (7)"},{"comment":"The EM derivation contains additional load-bearing errors. In the RTS-type smoother, Eq. (21) should read \\hat P_{t-1} = P^+_{t-1} + J_{t-1}(\\hat P_t - P^-_t)J^T_{t-1}, but the manuscript uses P^+_{t-1} instead of the predicted covariance P^-_t. More importantly, Eq. (23) uses \\hat P_t as if it were the second moment E[x_t x_t^T], whereas throughout the paper \\hat P_t denotes the smoothed covariance Cov[x_t|y]. The quadratic terms in (16) require second moments, so the M-step objective is missing the \\hat x_t \\hat x_t^T and \\hat x_{t-1} \\hat x_{t-1}^T contributions. The resulting EM parameter updates are therefore not maximum-likelihood updates for the stated model.","section":"III-B, Eqs. (19)-(23)"},{"comment":"The correction module is internally inconsistent. Equation (28) defines the update as xt = x^-_t + \\hat K_t(\\hat x_t - x^-_t) with \\hat K_t a graph filter, and Eq. (29) sets P^+_t = (I - \\hat K_t)P^-_t. But the justification preceding these equations, Eqs. (30)-(31), shows that the object being approximated is P^-_t H^T(HP^-_t H^T + sigma^2 I)^{-1} H, which is the product of the Kalman gain with H, not the Kalman gain itself, and it still contains the P^-_t factor. The covariance update in Eq. (32) is also different from Eq. (29), containing P^-_t \\hat K_t rather than \\hat K_t. Moreover, the claim that this can be represented as a graph filter is asserted 'up to the eigenspace mismatch between L and P^-_t' without any argument that P^-_t approximately commutes with L or is polynomial in L. Without such justification, the correction sub-module is not a principled unrolling of the Kalman correction step.","section":"IV-B, Eqs. (28)-(32)"}],"minor_comments":[{"comment":"The text states that sigma(·) is the sigmoid activation, but the update for z_t in Eq. (33) uses ReLU, and the observation-noise vector sigma_t from Eq. (25) shares the same symbol as the sigmoid function. Please clarify the intended activation and notation.","section":"IV-C, Eq. (33)"},{"comment":"The symbol Q is reused for both the expected negative log-likelihood and the system noise covariance Bdiag(alpha^2)B^T; different letters should be used to avoid confusion.","section":"III-B, Eq. (23)"},{"comment":"The text says 'codes are available here' but no URL or repository is provided; a working link is needed.","section":"V-A"},{"comment":"There are several typos, including 'for for' and 'becasue' in the water-network section and 'comlexity' in Section IV-C; these should be corrected.","section":"VI"},{"comment":"Although the standard deviations are reported as O(10^-3), Tables I-III show no error bars or confidence intervals; given the small performance differences among deep models, this makes it difficult to judge which differences are significant.","section":"V"}],"recommendation":"reject","confidential_remarks":"The errors in Eqs. (7), (19)-(23), and (28)-(32) are not localized; correcting them changes the model and requires re-running all experiments. The paper might be salvageable as a heuristic graph recurrent architecture, but the SPDE-based derivation and the EM/Kalman unrolling claims would need to be rebuilt. I therefore cannot recommend publication in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the reader and the stress-test have the load-bearing flaw right. Eq. (7) drops the identity term from the Euler step; the correct discrete transition is x_{t+1}=(I-cL)x_t+Bdiag(alpha)w_t, not -cLx_t+Bdiag(alpha)w_t. This is not a notational quibble. The missing identity propagates into the transition density, Kalman prediction (17), smoother equations (19)-(22), EM objective (23), training loss (35), and the water-network state equation (39). It also contradicts the paper's own Remark 1: the discrete model as written has mean evolution (-cL)^t mu0, which is not the heat-diffusion semigroup exp(-cLt)mu0. So the central claim that GKNet is a principled unrolling of a graph Kalman filter derived from the SPDE is not supported by the equations as they stand.\n\nThe paper is not without value. The architecture is a real new combination: joint learning of state and observation models, edge-wise uncertainty vector alpha, a recurrent inference module that generates the Kalman parameters, and a water-network case study with scarce sensors. The empirical results, though reported without error bars and with a placeholder code link, consistently show low-data gains over strong baselines. The EM and Bayesian framing is sensible apart from the missing identity.\n\nThe flaw is fixable in principle: replace -cL by I-cL wherever it appears, and rerun the EM, Kalman, and GKNet experiments. But as written, the experiments evaluate a different dynamical system from the one claimed. The authors also need to report error bars or seed-level results and make the code accessible.\n\nThis is a paper for model-based deep learning researchers working on graph time series, particularly those concerned with scarce data and uncertain topology. I would not desk reject it: the idea is substantive, the error is concrete and identifiable, and a serious referee can force the correction. But the corrected derivation and rerun experiments must be a condition of acceptance.\n\nReading group? Maybe—the failure mode is instructive. I would not cite it in its current form.","headline":"A useful architecture and honest experiments, but the central discretization is wrong and contradicts the paper's own Remark 1; send to referees with a demand to fix and rerun.","tokens_in":21926,"tokens_out":3764,"would_cite":false,"duration_ms":40303,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes a graph state-space model in which a latent signal diffuses over the graph with edge-level noise and is observed through a graph filter, and it argues that both the dynamics and the observation model can be learned…","keywords":["graph signal processing","state space models","graph Kalman filter","model-based deep learning","time series forecasting","graph neural networks","data efficiency","uncertain graphs"],"falsifier":"Simulate a known diffusion $dx_t = -cLx_t dt + B\\operatorname{diag}(\\alpha)d\\beta_t$ on a small graph, fit the proposed EM and GKNet to one-step transitions, and compare the estimated transition matrix against both $I - cL$ and $-cL$; the correct discrete model is the one that matches the simulated conditional mean $E[x_{t+1}|x_t]$, and a mismatch would falsify the state equation used in Equations (7), (17), and (35).","tokens_in":20697,"feed_emoji":"📈","tokens_out":7519,"duration_ms":78584,"temperature":0.7,"pith_summary":"The paper is trying to establish that a graph time series can be modeled as a latent diffusion process with edge-level noise, observed through a graph filter, and that both the dynamics and the observation model can be learned from partial measurements. The authors first solve this with an expectation-maximization algorithm wrapped around a graph Kalman smoother, then unroll the same recursions into a deep architecture, GKNet, that replaces the analytic Kalman gain with learned graph filters driven by a recurrent inference module. The payoff would be a forecasting and imputation method that stays accurate when training data are scarce and when the graph is only approximately known, as the stormwater case study argues. If right, this gives a principled way to inject graph structure into deep time-series models without requiring large training sets.","feed_headline":"Graph Kalman net learns dynamics from scarce sensor data","feed_subtitle":"It learns graph dynamics and observation models jointly, staying accurate with few sensors and few training samples.","key_machinery":"The load-bearing object is the graph state-space model, whose state transition is the graph Laplacian $L$ acting as a diffusion operator, whose noise is dispersed along edges through the node-to-edge incidence matrix $B$ with learnable per-edge strengths $\\alpha$, giving covariance $Q = B\\operatorname{diag}(\\alpha^2)B^\\top$, and whose observation operator is a masked graph filter $H = M\\tilde{H}(L)$ of order $K$. In GKNet, the same structure is unrolled: the prediction step uses $-cL$ and $Q$, the correction step uses a graph filter $\\hat{K}_t(L)$ in place of the Kalman gain, and a recurrent inference module outputs the filter coefficients, diffusivity $c$, and time-varying edge uncertainties from encoder statistics.","core_discovery":"The central claim is that the pair consisting of a graph-diffusion state equation and a graph-filtered observation equation is identifiable and learnable for graph time series. The state evolves as a stochastic heat diffusion $dx_t = -cLx_t dt + B\\operatorname{diag}(\\alpha)d\\beta_t$, with per-edge uncertainty $\\alpha$; observations are $y_t = M\\tilde{H}(L)x_t + v_t$, a sampled graph-filtered version of the state. The paper argues that this parameterization has few parameters, that the maximum-likelihood and EM route is the principled baseline, and that unrolling the graph Kalman recursions in GKNet, replacing the Kalman gain by a learned graph filter and the model parameters by an RNN hidden state, yields a scalable architecture that learns state and observation parameters jointly and end to end. In their experiments, the claim is that this design keeps forecasting and interpolation errors low when training data are scarce and the graph is noisy.","pith_inferences":["The per-edge uncertainties suggest a direct diagnostic use: train the model on a graph whose edges have been artificially perturbed and check whether the learned $\\alpha$ values concentrate on those edges, which would validate the uncertainty interpretation.","Because the Kalman-gain filter order is independent of graph size, the same recursions could be applied to a differently sized graph after retraining the encoder and decoder; the water-network transfer experiment is a partial demonstration.","The runoff-driven extension indicates the framework can handle exogenous graph inputs, so a natural next step is multi-step ahead forecasting by feeding predicted states back through the learned input filter."],"forward_implications":["If the state-space model is correctly learned, missing node values can be imputed from partial observations because the diffusion and the graph filter constrain what an unobserved node should do.","When training data are limited, the model's inductive bias, a graph diffusion plus edge-local noise, lets it forecast and interpolate in regimes where fully data-driven neural networks diverge.","The learned edge uncertainties $\\alpha$ provide a per-edge measure of where the assumed graph differs from the true process, so the same framework can flag unreliable connections.","Because GKNet's inference module has size governed by the filter order rather than the node count, the Kalman-gain and uncertainty filters scale to larger graphs than the $O(TN^3)$ EM route."],"supporting_citations":[{"why":"Supplies the SPDE graph kernel whose recursive state-space form anchors the proposed state model.","marker":"[25]"},{"why":"Provides the neural-aided graph Kalman tracking baseline that motivates replacing the analytic gain with a learned filter.","marker":"[39]"},{"why":"Establishes the Kalman neural network approach that the proposed architecture extends to graph settings.","marker":"[37]"},{"why":"Provides the expectation-maximization machinery used for likelihood inference with latent states.","marker":"[46]"},{"why":"Provides the Kalman filtering and smoothing recursions used in the E-step.","marker":"[49]"},{"why":"Defines graph filters and graph convolutional layers used in the observation model, encoder, and decoder.","marker":"[12]"},{"why":"Supplies the traffic benchmark datasets and a convolutional recurrent baseline used in the forecasting experiments.","marker":"[52]"},{"why":"Supplies the stormwater network dataset and the graph neural network baseline used in the water-network case study.","marker":"[58]"}],"fun_headline_variants":["Graph Kalman net learns sparse graph time series","Deep graph Kalman filter for noisy, sparse data","GKNet: joint state and model learning on graphs","Neural Kalman filtering for graph time series","Graph-aware deep Kalman for scarce sensor data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the discrete transition $x_{t+1} = -cLx_t + B\\operatorname{diag}(\\alpha)w_t$ being a faithful Euler discretization of the graph diffusion; if the step should read $x_{t+1} = (I - cL)x_t + \\cdots$, then the Kalman prediction, likelihood, and loss all change.","fun_headline_variants_meta":{"raw":{"variants":["Graph Kalman net learns sparse graph time series","Deep graph Kalman filter for noisy, sparse data","GKNet: joint state and model learning on graphs","Neural Kalman filtering for graph time series","Graph-aware deep Kalman for scarce sensor data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000149,"raw_usage":{"total_tokens":1215,"prompt_tokens":986,"completion_tokens":229,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":156}},"tokens_in":602,"tokens_out":229,"duration_ms":3139,"temperature":1.0,"reasoning_tokens":156,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:14:22.911931+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a known diffusion $dx_t = -cLx_t dt + B\\operatorname{diag}(\\alpha)d\\beta_t$ on a small graph, fit the proposed EM and GKNet to one-step transitions, and compare the estimated transition matrix against both $I - cL$ and $-cL$; the correct discrete model is the one that matches the simulated conditional mean $E[x_{t+1}|x_t]$, and a mismatch would falsify the state equation used in Equations (7), (17), and (35).","supporting_citations":[{"cited_title":"Non-separable spatio- temporal graph kernels via spdes,","cited_arxiv_id":null,"evidence_quote":"Supplies the SPDE graph kernel whose recursive state-space form anchors the proposed state model."},{"cited_title":"GSP-KalmanNet: Tracking Graph Signals via Neural-Aided Kalman Filtering","cited_arxiv_id":"2311.16602","evidence_quote":"Provides the neural-aided graph Kalman tracking baseline that motivates replacing the analytic gain with a learned filter."},{"cited_title":"Kalmannet: Neural network aided kalman filtering for partially known dynamics,","cited_arxiv_id":null,"evidence_quote":"Establishes the Kalman neural network approach that the proposed architecture extends to graph settings."},{"cited_title":"New Results in Linear Filtering and Prediction Theory,","cited_arxiv_id":null,"evidence_quote":"Provides the Kalman filtering and smoothing recursions used in the E-step."},{"cited_title":"Graph filters for signal processing and machine learning on graphs,","cited_arxiv_id":null,"evidence_quote":"Defines graph filters and graph convolutional layers used in the observation model, encoder, and decoder."},{"cited_title":"Transferable and data efficient metamodeling of storm water system nodal depths using auto-regressive graph neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the stormwater network dataset and the graph neural network baseline used in the water-network case study."}],"review_version":1}