- 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
-
The paper addresses a meaningful limitation of standard flow matching by incorporating sensor dependencies into the probability path through graph structure.
-
The graph-spectral path is derived from a principled variational objective, yielding a closed-form solution with interpretable frequency-dependent interpolation.
-
The anomaly scoring mechanism coherently integrates velocity discrepancies across flow times and graph frequencies using theoretically motivated weights.
-
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
-
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.
-
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 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.
-
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
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.
| Model | SMAP | SMD | CICIDS | SWAN | Avg. |
|---|---|---|---|---|---|
| GDN | 0.1470 0.0165 | 0.2379 0.0065 | 0.0837 0.0255 | 0.6326 0.0186 | 0.2753 |
| GCAD | 0.1151 0.0236 | 0.2512 0.0121 | 0.2296 0.0605 | 0.5501 0.0088 | 0.2865 |
| CATCH | 0.1484 0.0008 | 0.2952 0.0008 | 0.0883 0.0006 | 0.4514 0.0010 | 0.2458 |
| GRASP | 0.2403 0.0110 | 0.2327 0.0128 | 0.1722 0.0174 | 0.6770 0.0122 | 0.3306 |
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
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
Dear Reviewer SQ36,
We appreciate your positive assessment and constructive feedback. We are pleased to provide detailed responses to address your concerns.
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.
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.
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.
The eigendecomposition of the graph Laplacian has complexity of O(N3). 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].
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.