acceptodds
Under review as a conference paper at ICLR 2027

Graph-Spectral Flow Matching for Multivariate Time Series Anomaly Detection

13 views

Abstract

Multivariate time series anomaly detection typically relies on evaluating discrepancies between observations and outputs produced by models trained on normal data. An alternative perspective is to characterize the distribution of normal data through the generative dynamics, i.e., the velocity field, of flow matching models. However, standard flow matching typically adopts linear probability paths that overlook dependencies among variables, leading to a misalignment with the structured data distribution. To address this issue, we propose GRASP, a flow matching framework with a graph-spectral path for multivariate time series anomaly detection. GRASP incorporates graph structure into the probability path by minimizing a fixed-endpoint action that combines kinetic energy with graph Dirichlet energy. This formulation yields a closed-form path based on graph-frequency-dependent hyperbolic interpolation. A velocity predictor trained on normal data then detects anomalies using weighted velocity discrepancies aggregated across source samples, flow times, and graph frequencies. Theoretically, we establish that GRASP is invariant to the choice of Laplacian eigenbasis and decompose its expected oracle anomaly score into bounded endpoint uncertainty and graph-frequency-weighted Fisher discrepancy. Experiments on four benchmarks demonstrate the superior anomaly detection performance of GRASP and validate the effectiveness of its graph-spectral path and weighting mechanism.

Primary area:
generative models
Keywords:
flow matching, graph machine learning, multivariate time series anomaly detection

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_SQ36OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes GRASP, a graph-spectral flow-matching framework for multivariate time-series anomaly detection. It derives a graph-regularized probability path with frequency-dependent interpolation and detects anomalies using weighted velocity discrepancies across flow times and graph frequencies. The method is theoretically analyzed and evaluated on four benchmark datasets.

Critical strengths
  1. The paper addresses a meaningful limitation of standard flow matching by incorporating sensor dependencies into the probability path through graph structure.

  2. The graph-spectral path is derived from a principled variational objective, yielding a closed-form solution with interpretable frequency-dependent interpolation.

  3. The anomaly scoring mechanism coherently integrates velocity discrepancies across flow times and graph frequencies using theoretically motivated weights.

  4. Experiments on four benchmarks show consistent improvements in PRC and Best-F1, supported by ablation studies, although GRASP does not achieve the best ROC on SMD.

Critical weaknesses
  1. The method relies on a fixed graph constructed from normal training data, which may not adequately capture nonlinear, or time-varying dependencies. This assumption may limit robustness when relationships among variables evolve.

  2. The graph-spectral formulation requires eigendecomposition or related preprocessing, which may become computationally expensive for high-dimensional time series. Further scalability evaluations would strengthen the practical case.

  3. The theoretical analysis focuses on an ideal population-optimal velocity predictor. How the theoretical score decomposition translates to a finite-capacity trained model remains insufficiently clear and warrants further empirical validation.

  4. Best-F1 uses a threshold selected with test labels, making it an oracle-style metric rather than a deployable evaluation. The results should emphasize threshold-free metrics and/or thresholds selected using validation data.

Questions for authors

See weaknesses.

3 comments

Dear Reviewer SQ36,

We appreciate your positive assessment and constructive feedback. We are pleased to provide detailed responses to address your concerns.

W1: The method relies on a fixed graph constructed from normal training data, which may not adequately capture nonlinear or time-varying dependencies. This assumption may limit robustness when relationships among variables evolve.

To evaluate the robustness of GRASP to the graph structure, we added two additional experiments.

First, we keep the trained models fixed and replace the graph at inference. Following the estimation-error experiment of [1], we re-estimate the graph using only the most recent 50%, 25%, and 10% of the training data, which leads to noisy and incomplete graphs. Using only 10% of the data results in relative changes (relative Frobenius distance between the symmetric normalized Laplacians) of 77.23%, 37.57%, 31.34%, and 45.04% on SMAP, SMD, CICIDS, and SWAN, respectively. The results are reported below. GRASP achieves almost the same performance with graphs estimated from 50%, 25%, and 10% of the training data as with the original graph.

GraphSMAPSMDCICIDSSWAN
PRCROCBest-F1PRCROCBest-F1PRCROCBest-F1PRCROCBest-F1
Original0.31490.70150.41780.48450.85670.50870.39960.81080.46150.75560.87580.6886
50% data0.31210.70530.41400.48910.85220.50970.39270.80420.46030.76970.88680.7040
25% data0.31530.70730.41600.48910.85210.50960.39390.80550.46030.76850.88520.7013
10% data0.31480.70630.41450.49010.85220.51020.39460.80740.45950.76440.88460.6987

Second, we conduct edge rewiring by removing a certain percentage of edges and adding the same number of edges in other locations. GRASP achieves almost the same performance on graphs with 10%, 20%, 40%, and 60% rewired edges as with the original graph.

GraphSMAPSMDCICIDSSWAN
PRCROCBest-F1PRCROCBest-F1PRCROCBest-F1PRCROCBest-F1
Original0.31490.70150.41780.48450.85670.50870.39960.81080.46150.75560.87580.6886
10% rewired0.31490.70080.41800.48380.85340.50820.39960.81040.46130.75400.87550.6887
20% rewired0.31800.70000.41910.48370.85390.50770.40170.81350.46240.75310.87390.6849
40% rewired0.31390.69920.41590.48420.85370.50810.40410.81550.46240.75830.87850.6894
60% rewired0.31090.69470.41310.48500.85420.50840.40620.81370.46210.75710.87740.6913

As shown in our ablation study, the Erdős–Rényi graph leads to a large performance drop. These results suggest that GRASP benefits from the graph structure but is robust to noisy and incomplete graphs. Since the path construction and weighting schedule of GRASP operate in the graph-frequency domain, it remains stable under perturbations that preserve the spectral structure relevant to detection.

W2: The graph-spectral formulation requires eigendecomposition or related preprocessing, which may become computationally expensive for high-dimensional time series. Further scalability evaluations would strengthen the practical case.

The eigendecomposition of the graph Laplacian has complexity of O(N3)\mathcal{O}(N^3). This is only computed once during training and reused during inference. For typical multivariate time series anomaly detection benchmarks with tens to hundreds of variables, this cost is negligible. For example, on SMAP and SMD, the training of GRASP takes only about 11.7 and 4.5 minutes on a single A100. For very large graphs, full eigendecomposition can be expensive. As we mentioned in the conclusion, this can be avoided with scalable approximations to graph spectral operations; e.g., we can approximate the spectral path with Chebyshev polynomials of the Laplacian, which only requires sparse matrix multiplications [2].

W3: The theoretical analysis focuses on an ideal population-optimal velocity predictor. How the theoretical score decomposition translates to a finite-capacity trained model remains insufficiently clear and warrants further empirical validation.

Theorem 2 characterizes that the expected score under the oracle velocity predictor identifies a graph-frequency-weighted discrepancy between normal and test path distributions. This, however, does not directly translate to a detection-performance guarantee for the trained model. The gap between the trained model and the oracle predictor can be measured by the loss. From the experiments (Table 1), we could say that the trained model is able to capture this graph-frequency-weighted discrepancy.

W4: Best-F1 uses a threshold selected with test labels, making it an oracle-style metric rather than a deployable evaluation. The results should emphasize threshold-free metrics and/or thresholds selected using validation data.

Best-F1 represents retrospective detection potential rather than performance at a deployable threshold, and we will emphasize more on the threshold-free ranking metrics, ROC and PRC.

To evaluate detection performance with a threshold selected solely from validation data, we follow the default setting in mTSBench [2] and set the threshold at the 92.5th percentile of normal validation scores. The table below compares GRASP with the strongest baselines, GDN, GCAD, and CATCH. GRASP achieves the best Val-F1 on SMAP and SWAN and the best average.

As shown in Table 1 in the main paper, on CICIDS, GRASP achieves the highest PRC and ROC but a lower Val-F1, indicating a distribution shift between validation and test scores. A better-calibrated threshold that adapts to such shifts could further improve the performance.

ModelSMAPSMDCICIDSSWANAvg.
GDN0.1470 ±\pm 0.01650.2379 ±\pm 0.00650.0837 ±\pm 0.02550.6326 ±\pm 0.01860.2753
GCAD0.1151 ±\pm 0.02360.2512 ±\pm 0.01210.2296 ±\pm 0.06050.5501 ±\pm 0.00880.2865
CATCH0.1484 ±\pm 0.00080.2952 ±\pm 0.00080.0883 ±\pm 0.00060.4514 ±\pm 0.00100.2458
GRASP0.2403 ±\pm 0.01100.2327 ±\pm 0.01280.1722 ±\pm 0.01740.6770 ±\pm 0.01220.3306

We hope our responses have addressed your concerns and would be happy to consider further suggestions.

[1] Stability properties of graph neural networks. IEEE Transactions on Signal Processing 2020.

[2] Convolutional neural networks on graphs with fast localized spectral filtering. NeurIPS 2016.

[3] mTSBench: Benchmarking Multivariate Time Series Anomaly Detection and Model Selection at Scale, TMLR 2026

AuthorsOpenReview ↗

Gentle Reminder

Thank you again for your time and constructive feedback. As the discussion period is coming to a close, we would like to follow up. We hope our answers addressed the concerns appropriately. If so, we kindly ask the reviewer for potential reconsideration of the score. We would also be happy to clarify any remaining concerns.

Best regards,

The Authors

Reviewer_qZ5KOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The authors present GRAph-SPectral Path (GRASP), a flow matching framework with a graph-spectral path for multivariate time series (MTS) anomaly detection. The idea of GRASP is to incorporate a graph's structure into the probability path by minimizing a fixed-endpoint action. The authors present experimental results on four datasets.

Critical strengths

S1: The paper provides theoretical underpinnings for their anomaly score that aggregates weighted velocity discrepancies across source samples, flow times, and graph frequencies. Overall the paper seems theoretically grounded with a number of proofs in the appendix.

S2: The proposed GRASP method performs very well, overtaking existing methods.

S3: The experimental results section compares against 13 competitive baselines.

Critical weaknesses

W1: (Not critical) The idea of flow matching for multivariate time series (MTS) anomaly detection is not fundamentally new. However, the graph-spectral probability path is novel in this paper.

W2: The paper, while readable, is very dense in theoretical and mathematical notations and I presume it would take most readers a good while to fully understand.

W3: The authors do not evaluate the time complexity or execution time of GRASP. They do mention under the limitations in the conclusions that the method may be time consuming.

Questions for authors

Q1: Could some execution time results be presented in the experimental section?

2 comments

Dear Reviewer qZ5K,

We appreciate your positive assessment and constructive feedback. We are pleased to provide detailed responses to address your concerns.

W1: (Not critical) The idea of flow matching for multivariate time series (MTS) anomaly detection is not fundamentally new. However, the graph-spectral probability path is novel in this paper.

We thank the reviewer for recognizing the novelty of the proposed graph-spectral path. We agree that flow matching itself is established, although no prior work has used it for MTS anomaly detection. Our contribution mainly lies in the graph-spectral probability path and the anomaly score derived from it. We will emphasize this positioning more clearly.

W2: The paper, while readable, is very dense in theoretical and mathematical notations, and I presume it would take most readers a good while to fully understand.

We will add an intuitive explanation after each main result and move some intermediate derivations to the appendix.

W3 and Q1: Could some execution time results be presented in the experimental section?

The inference efficiency evaluation is already provided in Appendix E.10. GRASP requires 38.4 seconds on SMAP and 31.7 seconds on SMD to score the full test sets. It is faster than A-Transformer, CATCH, and GCAD, but slower than lighter baselines, because the score aggregates across several flow times and source samples. For application scenarios where faster inference is preferred, GRASP also provides a controllable trade-off between inference efficiency and detection performance. As discussed in Appendix E.8, GRASP can be further accelerated by reducing the flow-time evaluation points and source samples with a bit of compromised performance. GRASP is also efficient in terms of training. On SMAP and SMD, the training takes about 11.7 and 4.5 minutes on a single A100. We will move this inference efficiency analysis to the main paper.

We hope our responses have addressed your concerns and would be happy to consider further suggestions.

AuthorsOpenReview ↗

Gentle Reminder

Thank you again for your time and constructive feedback. As the discussion period is coming to a close, we would like to follow up. We hope our answers addressed the concerns appropriately. If so, we kindly ask the reviewer for potential reconsideration of the score. We would also be happy to clarify any remaining concerns.

Best regards,

The Authors

Reviewer_YbV8OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GRASP, a graph-spectral flow matching framework for multivariate time-series anomaly detection. The method introduces a graph-informed probability path derived from kinetic and graph Dirichlet energies, together with a weighted velocity-discrepancy anomaly score. The authors provide theoretical analysis of the proposed path and scoring mechanism. Experiments on four benchmark datasets demonstrate competitive performance against existing anomaly detection methods.

Critical strengths
  1. The paper presents an interesting and well-motivated approach to incorporating graph structure directly into flow matching probability paths. The closed-form graph-spectral interpolation is a meaningful methodological contribution.

  2. The theoretical analysis is relatively comprehensive, and the experimental results are promising. GRASP achieves the best PRC and Best-F1 across all four datasets, with additional ablation and sensitivity studies supporting the proposed framework.

Critical weaknesses
  1. Limited independent gains from the graph-spectral path. The improvement of GRASP-mean over GRASP-lin is relatively small on several datasets, especially SMAP and SMD. It remains unclear how much of the final performance gain comes from the proposed path rather than the anomaly-score weighting.

  2. Gap between theory and practical performance. The theoretical decomposition relies on a population-optimal velocity predictor and does not directly establish improved anomaly detection for the learned model.

  3. Limited evaluation of graph robustness. The method relies on a fixed graph constructed from training data. Its sensitivity to inaccurate or changing graph structures is not sufficiently investigated.

Questions for authors
  1. Could the authors provide statistical significance analysis for the improvement of the graph-spectral path over the standard linear path, particularly on SMAP and SMD?

  2. Could the authors further separate the contributions of the graph-spectral path and the proposed weighting scheme, for example by evaluating the linear path with the same weighting mechanism?

  3. How robust is GRASP to noisy, incomplete, or dynamically changing graph structures?

  4. Since Best-F1 uses thresholds selected from test labels, could the authors also report detection performance using thresholds determined solely from validation data?

2 comments

Dear Reviewer YbV8,

We appreciate your positive assessment and constructive feedback. We are pleased to provide detailed responses to address your concerns.

W1 and Q1: Could the authors provide statistical significance analysis for the improvement of the graph-spectral path over the standard linear path, particularly on SMAP and SMD?

We run tt-tests comparing GRASP with its ablation variants. The table below reports the results on SMAP and SMD, with significance markers indicating differences relative to GRASP (∗p<0.05^{*}p<0.05, ∗∗p<0.01^{**}p<0.01.). The results show that GRASP is significantly better than other baselines in all the metrics.

W1 and Q2: Could the authors further separate the contributions of the graph-spectral path and the proposed weighting scheme, for example by evaluating the linear path with the same weighting mechanism?

We would like to emphasize that the weight mechanism is constructed based on the graph-spectral path. To separate the contributions of the probability path and the weights, we add a weighted-linear variant, GRASP-lin-w. The graph-informed variational problem recovers the linear path when τ=0\tau=0, and the corresponding weight ηk(t)\eta_k(t) becomes t2t^2. GRASP-lin-w therefore applies this t2t^2 weighting to GRASP-lin. The results are reported below. GRASP-lin-w generally improves over GRASP-lin, showing that the weighting schedule contributes positively. Comparing GRASP with GRASP-lin-w show that the best performance is achieved when both the graph-spectral path and weight mechanism are used, validating the complementary benefits of the graph-spectral path and the weight mechanism.

VariantSMAPSMD
PRCROCBest-F1PRCROCBest-F1
GRASP0.31490.70150.41780.48450.85670.5087
GRASP-mean0.2915$^{**}$0.6941‾∗∗\underline{0.6941}^{**}0.3920∗∗^{**}0.4801∗∗^{**}0.8385∗∗^{**}0.5005^{**}\
GRASP-lin-w0.3098‾∗\underline{0.3098}^{*}0.6852$^{**}$0.4076‾∗\underline{0.4076}^{*}0.4811‾∗\underline{0.4811}^{*}0.8482‾∗\underline{0.8482}^{*}\underline{0.5033}^{*}\
GRASP-ER0.2907$^{**}$0.6909∗∗^{**}0.3912∗∗^{**}0.4797∗∗^{**}0.8379∗∗^{**}0.4993^{**}\
GRASP-lin0.2910$^{**}$0.6879∗∗^{**}0.3911∗∗^{**}0.4762∗∗^{**}0.8366∗∗^{**}0.4968^{**}\

W2: Gap between theory and practical performance. The theoretical decomposition relies on a population-optimal velocity predictor and does not directly establish improved anomaly detection for the learned model.

Theorem 2 characterizes that the expected score under the oracle velocity predictor identifies a graph-frequency-weighted discrepancy between normal and test path distributions. This, however, does not directly translate to a detection-performance guarantee for the trained model. The gap between the trained model and the oracle predictor can be measured by the loss. From the experiments (Table 1), we could say that the trained model is able to capture this graph-frequency-weighted discrepancy.

W3 and Q3: How robust is GRASP to noisy, incomplete, or dynamically changing graph structures?

To evaluate the robustness of GRASP to the graph structure, we added two additional experiments.

First, we keep the trained models fixed and replace the graph at inference. Following the estimation-error experiment of [1], we re-estimate the graph using only the most recent 50%, 25%, and 10% of the training data, which leads to noisy and incomplete graphs. Using only 10% of the data results in relative changes (relative Frobenius distance between the symmetric normalized Laplacians) of 77.23%, 37.57%, 31.34%, and 45.04% on SMAP, SMD, CICIDS, and SWAN, respectively. The results are reported below. GRASP achieves almost the same performance with graphs estimated from 50%, 25%, and 10% of the training data as with the original graph.

GraphSMAPSMDCICIDSSWAN
PRCROCBest-F1PRCROCBest-F1PRCROCBest-F1PRCROCBest-F1
Original0.31490.70150.41780.48450.85670.50870.39960.81080.46150.75560.87580.6886
50% data0.31210.70530.41400.48910.85220.50970.39270.80420.46030.76970.88680.7040
25% data0.31530.70730.41600.48910.85210.50960.39390.80550.46030.76850.88520.7013
10% data0.31480.70630.41450.49010.85220.51020.39460.80740.45950.76440.88460.6987

Second, we conduct edge rewiring by removing a certain percentage of edges and adding the same number of edges in other locations. GRASP achieves almost the same performance on graphs with 10%, 20%, 40%, and 60% rewired edges as with the original graph.

GraphSMAPSMDCICIDSSWAN
PRCROCBest-F1PRCROCBest-F1PRCROCBest-F1PRCROCBest-F1
Original0.31490.70150.41780.48450.85670.50870.39960.81080.46150.75560.87580.6886
10% rewired0.31490.70080.41800.48380.85340.50820.39960.81040.46130.75400.87550.6887
20% rewired0.31800.70000.41910.48370.85390.50770.40170.81350.46240.75310.87390.6849
40% rewired0.31390.69920.41590.48420.85370.50810.40410.81550.46240.75830.87850.6894
60% rewired0.31090.69470.41310.48500.85420.50840.40620.81370.46210.75710.87740.6913

As shown in our ablation study, the Erdős–Rényi graph leads to a large performance drop. These results suggest that GRASP benefits from the graph structure but is robust to noisy and incomplete graphs. Since the path construction and weighting schedule of GRASP operate in the graph-frequency domain, it remains stable under perturbations that preserve the spectral structure relevant to detection.

Q4: Since Best-F1 uses thresholds selected from test labels, could the authors also report detection performance using thresholds determined solely from validation data?

Best-F1 represents retrospective detection potential rather than performance at a deployable threshold, and we will emphasize more on the threshold-free ranking metrics, ROC and PRC.

To evaluate detection performance with a threshold selected solely from validation data, we follow the default setting in mTSBench [2] and set the threshold at the 92.5th percentile of normal validation scores. The table below compares GRASP with the strongest baselines, GDN, GCAD, and CATCH. GRASP achieves the best Val-F1 on SMAP and SWAN and the best average.

As shown in Table 1 in the main paper, on CICIDS, GRASP achieves the highest PRC and ROC but a lower Val-F1, indicating a distribution shift between validation and test scores. A better-calibrated threshold that adapts to such shifts could further improve the performance.

ModelSMAPSMDCICIDSSWANAvg.
GDN0.1470 ±\pm 0.01650.2379 ±\pm 0.00650.0837 ±\pm 0.02550.6326 ±\pm 0.01860.2753
GCAD0.1151 ±\pm 0.02360.2512 ±\pm 0.01210.2296 ±\pm 0.06050.5501 ±\pm 0.00880.2865
CATCH0.1484 ±\pm 0.00080.2952 ±\pm 0.00080.0883 ±\pm 0.00060.4514 ±\pm 0.00100.2458
GRASP0.2403 ±\pm 0.01100.2327 ±\pm 0.01280.1722 ±\pm 0.01740.6770 ±\pm 0.01220.3306

We hope our responses have addressed your concerns. We will revise the paper accordingly and would be happy to consider further suggestions.

[1] Stability properties of graph neural networks. IEEE Transactions on Signal Processing 2020.

[2] mTSBench: Benchmarking Multivariate Time Series Anomaly Detection and Model Selection at Scale, TMLR 2026

Reviewer_NFwsOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a method to replace linear flow matching path with a closed-form, graph-frequency-dependent path to score anomalies using weighted velocity residuals. It reports gains on four benchmarks. While the idea is well motivated and supported by ablations, there are multiple weaknesses to the paper as follows. It is not clear if the gains are only due to the graph spectral path or some other source. In particular, the gaps between GRASP-lin, -ER and -mean are below the reported std. It seems like the improvements come from the weighting. The novelty is incremental as the method combines existing lagrangian flow and Flow Mismatching idea. Furthermore, the evaluation is insufficient. Specifically, many baselines are near chance, and SWaT, WADI, PSM or MSL are not used.

Critical strengths

Correct derivation along with a well-motivated idea Wide ablations, sensitivity and runtime analyses

Critical weaknesses
  • Gains may not be due to the graph-spectral path
  • Weak theory
  • Insufficient evaluation
  • Incremental novelty
Questions for authors

The paper might benefit from:

  • weighted-linear ablation
  • statistical significance tests
  • Stronger baseline methods and more widely used datasets
  • Justification for how \tau was selected.
2 comments

Dear Reviewer NFws,

Thank you for your constructive feedback. We are pleased to address your concerns in detail.

W1 and Q1: Weighted-linear ablation.

We would like to emphasize that the proposed weighting schedule is derived from the graph-spectral path. To separate the contributions of the path and the weighting schedule, we added a weighted-linear variant, GRASP-lin-w. The graph-informed variational problem recovers the linear path when τ=0\tau=0, and the corresponding weight ηk(t)\eta_k(t) becomes t2t^2. GRASP-lin-w therefore applies this t2t^2 weighting to GRASP-lin. The results are reported below. GRASP-lin-w generally improves over GRASP-lin, showing the contribution of time-dependent weighting. Comparing GRASP with GRASP-lin-w further evaluates the benefit of the graph-spectral path. Together, these comparisons support the complementary contributions of the path construction and weighting schedule.

VariantSMAPSMDCICIDSSWAN
PRCROCBest-F1PRCROCBest-F1PRCROCBest-F1PRCROCBest-F1
GRASP0.31490.70150.41780.48450.85670.50870.39960.81080.46150.75560.87580.6886
GRASP-mean0.29150.6941‾\underline{0.6941}0.39200.48010.83850.50050.37640.75930.44990.7552‾\underline{0.7552}0.8753‾\underline{0.8753}0.6881‾\underline{0.6881}
GRASP-lin-w0.3098‾\underline{0.3098}0.68520.4076‾\underline{0.4076}0.4811‾\underline{0.4811}0.8482‾\underline{0.8482}0.5033‾\underline{0.5033}0.3904‾\underline{0.3904}0.7912‾\underline{0.7912}0.4554‾\underline{0.4554}0.74630.87360.6827
GRASP-lin0.29100.68790.39110.47620.83660.49680.37400.75430.44930.74860.87140.6841

W2: Weak theory

GRASP is theoretically grounded: the graph-spectral path is derived from a fixed-endpoint variational problem as demonstrated in Theorem 1, and the weight design is driven by Lemma 1. Proposition 1 establishes the eigenbasis invariance of the model. Theorem 2 and Corollary 1 further show how the designed anomaly score captures the path distribution shift. All proofs are provided in the appendix. If the reviewer could point out specific points that they feel relevant, we'll be happy to analyze them.

Q2: statistical significance tests

We run tt-tests comparing GRASP with each baseline in Table 1. Among the 153 baseline--dataset--metric comparisons, GRASP has significantly higher performance in 150 comparisons (p<0.05p<0.05). The three exceptions are ROC on SMD, where CATCH, OmniAnomaly, and USAD perform significantly better.

W3 and Q3: Stronger baseline methods and more widely used datasets

We did not use the SWaT, WADI, PSM, and MSL datasets because they are actually criticized by previous benchmark studies [1, 2, 3]. These studies identify issues in these datasets, including distribution shifts, positional bias, potentially unlabeled anomalies, and anomalies dominated by a small number of high-magnitude channels. Therefore, we select alternative datasets from the recent mTSBench benchmark [3].

For the baselines, we included best-performing methods from the mTSBench benchmark [3] as well as several recent works. If the reviewer has additional methods in mind, we would be happy to consider them.

W4 : Incremental novelty

The main contribution of GRASP is the graph-spectral path and the anomaly score derived from it. Appendix A discusses the positioning of this work relative to prior work. We will make these distinctions more explicit in the main paper.

Q4: Justification for how τ\tau was selected.

The weight parameter τ\tau weights the graph Dirichlet energy against the kinetic term, setting the trade-off between fidelity and graph smoothness. This is the same trade-off underlying the unfolded GNN propagation rules [4, 5, 6]. A weight parameter around 2 is observed to perform well in these works. We therefore set it to 2 by default, though these two formulations are different. The results reported in Table 1 are obtained with this default value. To better understand the effect of τ\tau, we also performed a sensitivity analysis in Appendix E.9. The results show that the optimal τ\tau is different for different datasets, while the performance remains relatively stable for τ∈[0.5,4]\tau \in [0.5,4].

We hope our responses have addressed your concerns. We will revise the paper accordingly and would be happy to consider further suggestions.

[1] Current Time Series Anomaly Detection Benchmarks are Flawed and are Creating the Illusion of Progress, ICDE 2022

[2] TimeSeAD: Benchmarking Deep Multivariate Time-Series Anomaly Detection, TMLR 2023

[3] mTSBench: Benchmarking Multivariate Time Series Anomaly Detection and Model Selection at Scale, TMLR 2026

[4] A Unified View on Graph Neural Networks as Graph Signal Denoising, CIKM 2021

[5] Interpreting and Unifying Graph Neural Networks with An Optimization Framework, WWW 2021

[6] Implicit vs Unfolded Graph Neural Networks, JMLR 2025

AuthorsOpenReview ↗

Gentle Reminder

Thank you again for your time and constructive feedback. As the discussion period is coming to a close, we would like to follow up. We hope our answers addressed the concerns appropriately. If so, we kindly ask the reviewer for potential reconsideration of the score. We would also be happy to clarify any remaining concerns.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.