{"id":"23d17015-b749-49a7-92ee-3569bfa144b9","arxiv_id":"2501.13493","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Anomalies are detected by measuring deviations in dynamic, gradient-based Granger causality patterns between time series variables.","lead":"This paper presents GCAD, an anomaly detection method that builds a Granger causality graph from gradients of a deep predictor and scores anomalies by how much that graph deviates from normal. It reports state-of-the-art results on five public sensor datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5) defines the causality matrix from loss gradients, yet ∂L/∂x = 2(ŷ−y)∂ŷ/∂x vanishes when prediction error is zero, so the 'typical' causal pattern may encode prediction-error magnitude rather than Granger causal strength, leaving the central claim untested.","rationale":"The reader's weakest assumption identifies the loss-gradient measure as the fragile premise. My stress-test sharpens this: the premise is not merely empirically unverified; the derivative structure makes the conflation likely. Under Eq. (1), the gradient with respect to the input is proportional to the residual, so at the low training loss typical of normal windows, A_norm is attenuated and the ratio in Eq. (10) is inflated by any increase in prediction error. The paper provides no ablation separating prediction error from causal deviation, and the reported case study shows only that scores rise during an attack, which is also consistent with a residual detector. This does not overturn the empirical anomaly-detection results, but it does mean the central causal claim needs a decisive test before it can be accepted. Because the reader already conditioned acceptance on addressing exactly this issue, the verdict remains conditional; if the proposed Jacobian-based test shows no causal signal, the verdict would move toward reject.","tokens_in":11471,"tokens_out":6869,"duration_ms":69455,"concrete_test":"On SWaT, recompute the causality matrix with the predictor Jacobian ∂ŷ_{t,j}/∂x_{φ,i} substituted for ∂L_{t,j}/∂x_{φ,i} in Eq. (5), keeping sparsification and scoring (Eqs. 6–12) unchanged. Report (i) the Spearman correlation between GCAD's A entries and per-channel residuals |ŷ−y| on normal windows; (ii) AUROC/AUPRC with the Jacobian-based matrix. High correlation, or a large performance drop with the Jacobian, would show the current score is a prediction-error detector rather than a Granger-causality detector.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central causal quantity in Eq. (5) is a_{i,j} = ∫ |∂L_{t,j}/∂x_{φ,i}| P(x_{φ,i}) dx_{φ,i}, where L_{t,j} is the squared prediction error from Eq. (1). Because ∂L_{t,j}/∂x_{φ,i} = 2(ŷ_{t,j}−y_{t,j}) ∂ŷ_{t,j}/∂x_{φ,i}, the integrand is the product of the prediction residual and the input-Jacobian of the predictor. At any window where the trained predictor is accurate, the residual is near zero, so the Granger-causal effect is suppressed even if the predictor's output depends strongly on x_i. Moreover, the equality in Eq. (4) is only a first-order approximation: if L=0, then |L−L*| is second-order in the perturbation, so the limit is 0. Thus A_norm, computed on normal training windows with small residuals, is not a faithful estimate of the dependence structure in Definition 1; it is a residual-weighted Jacobian. Deviations in Eq. (10) can then be driven by changes in prediction error rather than by changes in causal patterns. This makes the central claim—that anomalies are detected through Granger-causal changes—unsubstantiated unless the loss-gradient construction is replaced or its causal content is validated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GCAD, an unsupervised anomaly detection method for multivariate time series. It first trains a TSMixer-style predictor with MSE loss on normal data. At test time, for each sliding window it back-propagates the per-channel squared error to compute gradients of each prediction channel with respect to each input channel and lag. These gradients are aggregated into a Granger causality matrix (Eq. 5), sparsified by removing symmetric contributions and low-magnitude entries (Eq. 6), and compared to a typical matrix estimated from Bernoulli-sampled training windows (Eqs. 7-8). The final anomaly score combines the causal deviation (Eq. 10) with a diagonal temporal deviation (Eq. 11). Experiments on SWaT, SMD, MSL, SMAP, and PSM report AUC/PRC against six baselines, an ablation study, and a SWaT case study.","tokens_in":11802,"tokens_out":6320,"duration_ms":58709,"significance":"If the gradient-based quantity in Eq. (5) were a faithful estimate of nonlinear Granger causality, the paper would contribute a useful interpretable dependency-detection mechanism: it avoids repeated optimization at test time, produces dynamic per-window graphs, and is evaluated on five standard benchmarks with a physically interpretable case study. The main weakness is that the causal quantity is defined through the prediction loss, making it a residual-weighted Jacobian rather than a pure dependence measure. The paper does not compare against the underlying prediction error, so the central interpretive claim is not yet empirically separated from a standard prediction-error detector. This is fixable but requires additional analysis and experiments.","major_comments":[{"comment":"The derivation from Definition 1 to Eq. (5) is not sound as written. Eq. (4) states that the limit of |L_{t,j}-L*_{t,j}| as Delta approaches 0 equals |partial L_{t,j}/partial x_{t',i}| |Delta|, but the left-hand side tends to 0 for any differentiable L; the correct statement is an asymptotic equivalence as Delta approaches 0. More substantively, since L_{t,j} = (y_hat_{t,j} - y_{t,j})^2, the derivative partial L_{t,j}/partial x_{t',i} equals 2(y_hat_{t,j} - y_{t,j}) times partial y_hat_{t,j}/partial x_{t',i}. The quantity a_{i,j} in Eq. (5) is therefore a residual-weighted Jacobian, not a property of the predictor's input-output map f alone. On normal training windows with small residuals, a_{i,j} under-weights genuinely strong dependencies, and on anomalous windows the score in Eq. (10) can increase simply because prediction error increases. The paper therefore does not currently establish that GCAD detects changes in Granger causal structure; it establishes only that it detects a transformed prediction-error signal. Please replace L by y_hat in the gradient definition, or otherwise provide validation that the residual weighting does not drive the reported results.","section":"Granger Causality Discovery, Eqs. (4)-(5)"},{"comment":"No experiment compares GCAD to the plain prediction-error score of the same TSMixer predictor. Since the gradient in Eq. (5) contains the prediction residual as a factor, the relative improvement over a thresholded MSE baseline is necessary to support the causal-deviation claim. The ablation '-GC' in Table 3 is not defined in the text, so it is unclear whether it removes the gradient transformation or replaces the score with raw prediction error. Without such an ablation, the competitive AUC/PRC results in Table 2 are consistent with the hypothesis that GCAD is a prediction-error detector with an additional nonlinear transformation.","section":"Experimental Results, Tables 2 and 3"},{"comment":"The anomaly score in Eq. (10) compares test causal matrices to a pointwise training-set average A_norm sampled with Bernoulli(p) in Eq. (7). Because A_norm is an average over normal windows, a distribution shift in the test input that changes the Jacobian but not the causal relation will still produce a large score; likewise, a true causal change that happens to preserve the residual-weighted gradient magnitude will be missed. The paper should include an experiment that distinguishes these cases, for example by evaluating on windows with synthetic additive noise but unchanged causal structure, or by reporting the score decomposition into residual and Jacobian components.","section":"Causal Deviation Scoring, Eqs. (7)-(12)"},{"comment":"Eq. (5) defines a_{i,j} as an integral with respect to a distribution P over x_{phi,i}, with limits of integration t-tau to t-1, but the experiments appear to compute pointwise gradients on each observed window G_t (Figure 1 and the 'Prediction-based Gradient Generator' section). If the implementation uses pointwise gradients, Eq. (5) is not what is evaluated; if it uses an empirical average, the Monte Carlo estimator should be stated. Either way, the uniform-distribution simplification is undefined over the unbounded support of the input variables, and the integral notation with time-lag limits and an input-space integration variable is confusing.","section":"Granger Causality Discovery, Eq. (5) and implementation"}],"minor_comments":[{"comment":"The implementation details omit concrete values for the hyperparameters tau, h, beta, p, and epsilon; these should be reported for reproducibility, along with the predictor depth and width.","section":"Experimental Setup"},{"comment":"The notation W'_train = W_train ⊙ B mixes a time-indexed matrix with a vector of length n_train; please clarify the dimensions and the exact Bernoulli sampling procedure.","section":"Causal Deviation Scoring, Eq. (7)"},{"comment":"Averages over 10 runs are reported without standard deviations or significance tests, so small differences such as the SWaT AUROC gap of 0.8690 versus 0.8493 may be within run-to-run noise.","section":"Experimental Results, Table 2"},{"comment":"There are several rendering issues, including 'by_t = f' instead of a hat notation, and the matrix entries in Eq. (6) are printed as 'eAi,j' and 'Ai,j - AT i,j'; please fix these formatting problems.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper's core contribution depends on interpreting the gradient quantity as Granger causality, but the current formulation is a residual-weighted Jacobian and the experiments do not separate it from a plain prediction-error baseline. I recommend major revision rather than rejection because the idea is novel and the missing validation is a standard addition; however, if the authors cannot demonstrate that the causal transformation adds information beyond prediction error, the contribution may be too incremental for this venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: GCAD is a new and plausible way to do multivariate time-series anomaly detection—train a predictor, derive a per-pair \"causal\" matrix from gradients of the per-channel loss, sparsify, and score test windows by deviation from the training average. The combination is genuinely novel, and the results on SWaT, SMD, MSL, SMAP, and PSM are strong: AUROC between 0.87 and 0.95 against six baselines. The SWaT case study is a nice touch; the sensors flagged as causally deviant match the physical attack path.\n\nWhat the paper does well: it avoids online optimization at test time, the transpose-subtraction sparsification is simple and effective in the ablation, and the authors are honest about the limitation to pairwise causal relations.\n\nThe soft spots sit in the load-bearing wall. Equation (5) defines a_{i,j} as an integral of |∂L/∂x| over a distribution P, but the implementation, as far as the text shows, computes the pointwise gradient at the observed window. No integral or distribution appears in the algorithm. More importantly, because L is the squared prediction error, ∂L/∂x = 2(ŷ−y)·∂ŷ/∂x. On normal training windows where the predictor is accurate, the residual is near zero, so the \"typical\" causal matrix is a residual-weighted Jacobian. Deviations in Eq (10) could then reflect changes in prediction-error magnitude rather than changes in causal pattern. The stress-test note lands. This does not kill the method as an anomaly detector—it may be a fancier prediction-error detector—but it does mean the central interpretability claim is unsubstantiated as written.\n\nThe missing baseline makes this worse: the experiments never compare against plain prediction error on the same predictor, with or without the sparsification. Without that, we cannot tell whether the causal scoring adds anything beyond the predictor's own error signal. Hyperparameter reporting is also thin: τ and h are varied only for SWaT; the other four datasets don't report settings, and p, β, and ε are not analyzed at all.\n\nIf I were editing, I would not desk reject. The combination is new, the empirical work is substantial, and the issues are addressable. Send it to referees with a request for (1) a prediction-error baseline, (2) an implemented version of Eq (5) or an explicit statement that the pointwise gradient is the intended quantity, and (3) full hyperparameter reporting. As it stands, I would treat the causal interpretation as a hypothesis, not a demonstrated fact.\n\nFor a reading group, it is a good case study in how gradient-based attribution claims can over-read. I would maybe bring it up, but I would not cite it in my own work until the confound is addressed.\n\nTake care.","headline":"GCAD is a genuinely novel gradient-based Granger-causality scoring method with strong benchmark results, but the causal quantity it defines is not what it computes, and the missing prediction-error baseline leaves the core interpretability claim unsubstantiated.","tokens_in":12271,"tokens_out":2541,"would_cite":false,"duration_ms":22717,"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":"GCAD detects multivariate time series anomalies by measuring deviations in Granger causality between sensors, computed from the gradients of a deep predictor.","keywords":["anomaly detection","multivariate time series","Granger causality","gradient-based causal discovery","dynamic causal graph","deep learning interpretability","causal deviation scoring","sensor networks"],"falsifier":"Run GCAD on a synthetic system with known nonlinear causal structure, and perturb the test data in two ways: break one causal edge (a true anomaly) and scale the marginal variance of all sensors without touching any causal edge (a benign distribution shift). If the causal deviation score does not rise sharply for the edge break and stay flat for the variance shift, the central premise is falsified.","tokens_in":11300,"feed_emoji":"📈","tokens_out":8631,"duration_ms":67129,"temperature":0.7,"pith_summary":"The paper tries to show that multivariate time series anomalies are visible as changes in Granger causality between sensors, and that these causal changes can be read off from the gradients of a trained neural predictor. Concretely, GCAD trains a simple all-MLP predictor on normal data, then for each sliding window computes a Granger causality matrix whose entry $a_{i,j}$ is the accumulated magnitude of the gradient of sensor $j$'s prediction loss with respect to the past values of sensor $i$. After sparsifying away symmetric similarity edges, the method compares each test window's causal matrix with the typical normal matrix; large relative deviations flag anomalies. If the premise holds, this turns anomaly detection into monitoring an interpretable, dynamic dependency graph rather than raw prediction errors.","feed_headline":"Anomalies appear as breaks in sensor-to-sensor causality","feed_subtitle":"GCAD reads Granger causality from a predictor's gradients and scores anomalies as graph deviations.","key_machinery":"The load-bearing object is the channel-separated gradient tensor $G_t \\in \\mathbb{R}^{N\\times N\\times \\tau}$: for each target channel $j$, the gradient of the squared prediction error with respect to every past input value of every source channel $i$. Integrating absolute gradient magnitudes over the time lag (Eq. 5) turns this tensor into a Granger causality matrix $A$, whose off-diagonal entries encode directed spatial dependence and whose diagonal entries encode temporal self-dependence. Two further mechanisms carry the argument: a symmetry-based sparsification $\\tilde{A}_{i,j} = \\max(0, A_{i,j} - A_{j,i})$ that strips bidirectional similarity, and the relative deviation score in Eq. (10) that separates anomalous causal patterns from the normal baseline. Because the predictor is fixed after training, all of this is computed in one forward plus backward pass per window, with no online optimization.","core_discovery":"The central claim is that an anomaly in a multivariate sensor stream is, at root, a change in the pairwise Granger causal pattern, and that the pattern can be measured cheaply at test time by backpropagating each channel's squared error through a pretrained predictor. In the paper's setup, the Granger causality matrix $A$ has entries $a_{i,j} = \\int_{t-\\tau}^{t-1} |\\partial L_{t,j}/\\partial x_{\\phi,i}|\\, dP(x_{\\phi,i})$, where $L_{t,j}$ is the prediction error for channel $j$ and $\\phi$ indexes past time steps; this makes the causality effect proportional to how much sensor $i$'s past values influence the predictor's error on sensor $j$. The matrix is sparsified by subtracting its transpose (keeping only unidirectional edges) and thresholding, and the anomaly score is the relative $\\ell^1$ deviation of the test matrix from the averaged normal matrix, plus a weighted term for diagonal (temporal self-dependence) deviations. On SWaT, SMD, MSL, SMAP, and PSM, GCAD reports higher AUROC and AUPRC than six prior methods on most metrics.","pith_inferences":["The same gradient-based causality estimator could be applied to any differentiable predictor, so the core idea generalizes beyond the specific all-MLP network used here; any model whose output is differentiable with respect to its inputs can yield a causal graph.","The per-pair deviation matrix $D$ could be thresholded per edge to produce a causal 'diff' map, converting a scalar anomaly score into a root-cause localization tool for operators.","The paper's success on SWaT, where single-sensor readings barely change during attacks, suggests that interaction-structure changes are an earlier and more reliable anomaly signal than marginal distribution shifts in real control systems.","Since the authors note only pairwise causal effects are captured, a natural extension is to compute gradients of joint or conditional losses over groups of channels, which might expose anomalies that only appear in coordinated multi-sensor behavior."],"forward_implications":["Anomaly detection becomes interpretable: the deviation matrix $D = |\\tilde{A}_{test} - A_{norm}|$ shows exactly which sensor pairs' causal links have changed, enabling diagnosis rather than just flagging.","The method works on streaming data without retraining or parameter adjustment during testing, since causality is extracted from gradients of the already-trained predictor.","Because the causal matrix encodes both off-diagonal spatial effects and diagonal temporal effects, a single pipeline captures both types of dependency change.","On five real-world benchmarks, the causal-pattern deviation score outperforms prediction-error and reconstruction baselines in AUROC and AUPRC on most datasets.","Sparsifying the causal graph improves accuracy, indicating that removing bidirectional similarity noise strengthens the causal signal."],"supporting_citations":[{"why":"Supplies the original definition of Granger causality that the paper generalizes to nonlinear deep predictors.","marker":"Granger 1969"},{"why":"Provides the definition of nonlinear Granger causality (Definition 1) that the gradient-based quantification is designed to match.","marker":"Cheng et al. 2022"},{"why":"The TSMixer all-MLP architecture used as the prediction-based gradient generator; its differentiability enables the gradient extraction.","marker":"Chen et al. 2023"},{"why":"GDN is the primary graph-based baseline that motivates the need for interpretable, dynamic dependency modeling.","marker":"Deng and Hooi 2021"},{"why":"Contrast method: discovers Granger causality from sparse penalized network weights, which the paper argues is unsuitable for streaming anomaly detection.","marker":"Tank et al. 2021"},{"why":"Provides the SWaT water-treatment dataset used for evaluation and case study.","marker":"Mathur and Tippenhauer 2016"},{"why":"Provides the SMD dataset and a baseline (LSTM-VAE) for comparison.","marker":"Su et al. 2019"},{"why":"Provides the MSL and SMAP spacecraft datasets used in evaluation.","marker":"Hundman et al. 2018"}],"fun_headline_variants":["Causality shifts expose anomalies in sensor data","GCAD: anomalies are breaks in Granger causality","When sensor causality breaks, it's an anomaly","Detect anomalies via Granger graph changes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the influence one sensor's past values has on the model's error for another sensor, as measured through the trained predictor's gradients, faithfully reflects true causal influence, so that a change in those gradient magnitudes means an anomaly rather than just a shift in the data.","fun_headline_variants_meta":{"raw":{"variants":["Causality shifts expose anomalies in sensor data","GCAD: anomalies are breaks in Granger causality","When sensor causality breaks, it's an anomaly","Detect anomalies via Granger graph changes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000306,"raw_usage":{"total_tokens":1751,"prompt_tokens":940,"completion_tokens":811,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":752}},"tokens_in":556,"tokens_out":811,"duration_ms":7283,"temperature":1.0,"reasoning_tokens":752,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:53:02.882601+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GCAD on a synthetic system with known nonlinear causal structure, and perturb the test data in two ways: break one causal edge (a true anomaly) and scale the marginal variance of all sensors without touching any causal edge (a benign distribution shift). If the causal deviation score does not rise sharply for the edge break and stay flat for the variance shift, the central premise is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the original definition of Granger causality that the paper generalizes to nonlinear deep predictors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the definition of nonlinear Granger causality (Definition 1) that the gradient-based quantification is designed to match."},{"cited_title":"O.; Yoder, N","cited_arxiv_id":null,"evidence_quote":"The TSMixer all-MLP architecture used as the prediction-based gradient generator; its differentiability enables the gradient extraction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GDN is the primary graph-based baseline that motivates the need for interpretable, dynamic dependency modeling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contrast method: discovers Granger causality from sparse penalized network weights, which the paper argues is unsuitable for streaming anomaly detection."}],"review_version":1}