acceptodds
Withdrawn from ICLR 2027

Explainable Multivariate Time-Series Anomaly Detection via Dynamic Graph Optimal Transport

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multivariate time-series anomaly detection is challenging because anomalies often arise from changes in cross-sensor dependencies rather than large deviations in individual channels. Existing methods capture amplitude errors but often miss these structural anomalies and offer limited support for explainable sensor-level root-cause localization. We propose TOAD, an explainable topology-aware optimal-transport framework for unsupervised multivariate time-series anomaly detection and root-cause attribution. TOAD constructs dynamic sensor graphs, learns topology-aware spatiotemporal latent representations, and detects structural anomalies by comparing latent distributions with regime-aware normal references using sliced Wasserstein distance. To improve robustness, the final anomaly score integrates evidence from transport, reconstruction, and temporal transitions. Beyond detection, TOAD provides faithful sensor-level explanations by attributing anomaly scores through gradient sensitivity, counterfactual score reduction, and graph-aware propagation, enabling root-cause localization directly aligned with the detection mechanism. Extensive experiments on five public benchmarks demonstrate that TOAD achieves the best threshold-independent detection performance on four datasets and improves root-cause localization from 64.0% to 72.8% (H@100) and from 78.2% to 86.4% (H@150), providing accurate and actionable explanations for anomalous events.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate time-series, Explainable Anomaly Detection, Dynamic Graph Learning, Optimal Transport, Unsupervised Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Xof2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes TOAD, a method for unsupervised anomaly detection and root-cause analysis in multivariate time series. For each window, it builds a Pearson correlation graph, encodes the window with a GCN and a temporal convolution, and pools everything into one embedding. The anomaly score combines three terms: a sliced Wasserstein distance between recent embeddings and regime-specific normal references, a reconstruction error, and a temporal-transition term. Sensor-level root causes are obtained from the gradient of this fused score, so the explanation is tied to the detector's own decision. The authors also give Lipschitz and concentration results that claim stable scores and stable attributions. Experiments on SMD, SMAP, MSL, PSM and SWaT show the best VUS-ROC on four datasets, better hit rates on synthetic sensor injections in SMD, and a counterfactual repair test for faithfulness.

Critical strengths
  • The paper targets a real gap: anomalies that change cross-sensor dependencies while each sensor stays in its normal range. These are hard to catch with pointwise reconstruction or forecasting errors and deserve more attention.
  • Tying the explanation to the same score that raises the alarm is the right idea for operators. The paper also keeps faithfulness separate from causal correctness, which is a fair and useful distinction.
  • Regime-specific references with a minimum over regimes is a simple and practical way to avoid false alarms when the system moves between normal operating modes. Many detectors ignore this.
  • Fusing transport, reconstruction and temporal evidence, with robust normalization fitted on training data only, is a clean design. It keeps each stream on a comparable scale and avoids using test statistics for normalization.
  • Using counterfactual repair to test whether attributed sensors actually drive the score is a good evaluation idea in principle. It checks the explanation against the model's decision instead of only against proxy labels.
Critical weaknesses
  • The RCA method is not clearly defined. TOAD's own gradient attribution (21.3 in Table 5) is worse than GDN (23.5) and TranAD (25.8) at reducing TOAD's own score, so the gain of "TOAD RCA" comes from steps that are never described.
  • The theory does not match the model. The proofs assume a fixed graph, independent windows and smooth activations, but TOAD recomputes the graph per window, uses overlapping stride-1 windows and LeakyReLU.
  • Some proofs are wrong even under their own assumptions. Theorem 1 says the false alarm bound improves with d and T, but X_max bounds the Frobenius norm of a window and grows with sqrt(dT) for bounded sensor values, so L_f · X_max = K · x̄ and the d and T terms cancel. Theorem 2(iii) bounds the probability over the random draw of training windows while the test input is fixed. So it only says attributions concentrate around their mean, not that a wrong sensor is unlikely to be flagged during deployment.
  • Hyperparameters seem to be tuned on test data. The chosen values in Figure 4 are at the peak for all five datasets, and no validation split is mentioned.
  • Some baseline numbers look odd. All Table 1 means for DADA, GDN, KAN-AD and KMeansAD are exact integers, and PCA on SMD has 83.1 AUC-ROC but 39.9 VUS-ROC.
  • RCA is tested only on synthetic data, although real labels exist. SMD has interpretation labels, used by OmniAnomaly and TranAD to report H@P%, and the SWaT attack list names the attacked sensors.
  • The value of OT is not shown. With stride 1 and W=10 the windows almost fully overlap, and W=1 loses only about 1.5 points, so a kNN baseline on the same embeddings and a comparison with MADGA are needed.
  • Gains are small on VUS-PR, the metric TSB-AD recommends. On SMD (+0.2) and SMAP (+0.3) they are within one standard deviation.
  • The method is described inconsistently. Figure 1 and Eq. 6 list different losses, and L_reg is always zero after hard projection, yet removing it changes results.
  • The dataset claims are weak. SMAP and MSL are mostly binary command channels where Pearson graphs are undefined, and SWaT attacks are not mainly slow single-sensor drifts.
Questions for authors
  • What exactly is "TOAD RCA" in Tables 4 and 5? Please give the full algorithm, including the counterfactual term and graph-aware propagation, since Sec. 3 only defines a gradient norm.
  • Why does TOAD's own gradient attribution reduce TOAD's score less than GDN and TranAD rankings in Table 5?
  • Can the theorems be corrected for a per-window graph, overlapping windows and LeakyReLU? Which guarantees still hold for the evaluated model?
  • Does the counterfactual step inside TOAD RCA use the same repair operator as the faithfulness test? If yes, how is the test not circular?
  • How were T, k, R, W and the fusion weights chosen? Was Figure 4 computed on test data, and if so, what are the results with a separate validation split?
  • Can you share per-seed results or logs for the baselines in Table 1? Why are all means for DADA, GDN, KAN-AD and KMeansAD exact integers, and why does PCA on SMD have 83.1 AUC-ROC but 39.9 VUS-ROC?
  • Why was RCA not evaluated on the real SMD interpretation labels or the SWaT attack list? What are H@P% and N@P% on these?
  • Which training objective produced the reported results, the one in Figure 1 or Eq. 6?
  • With hard spectral projection after every step, L_reg is always zero. What does "w/o spectral reg." actually remove?
  • How are constant channels handled in the Pearson graph, especially on SMAP and MSL where most channels are binary command flags?
  • In the faithfulness test, windows are picked by TOAD's score and all methods are judged under TOAD's threshold. How do the results change if windows are picked by ground-truth labels?
  • TSB-AD recommends VUS-PR as the most reliable metric, so why did you rank methods by VUS-ROC? Under VUS-PR (Table 6), the gains on SMD (+0.2) and SMAP (+0.3) are within one standard deviation. Do you still claim state-of-the-art detection on these datasets?

Suggestions

  • Please compare against a kNN or Mahalanobis score on the same embeddings, and against MADGA, to show what OT adds.
Reviewer_s2jAOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes TOAD, an unsupervised framework for multivariate time-series anomaly detection and sensor-level root-cause localization. TOAD constructs dynamic sensor graphs and learns topology-aware spatiotemporal representations, using sliced Wasserstein distance to detect deviations from normal latent distributions. It further combines optimal transport, reconstruction, and temporal-transition signals for anomaly scoring, and employs gradient-based attribution to identify influential sensors. Experiments on five public benchmarks show competitive anomaly detection performance, with the best VUS-ROC on four datasets, along with improved sensor-level localization and counterfactual faithfulness results.

Critical strengths

1. Practical Relevance. The paper addresses the important problem of multivariate time-series anomaly detection and sensor-level root-cause localization within a unified framework, offering potential practical value for system monitoring and diagnosis.

2. Comprehensive Experimental Evaluation. The paper evaluates TOAD on five public benchmarks using multiple detection metrics, ablation studies, and counterfactual faithfulness tests. The results demonstrate competitive detection performance and promising sensor-level localization accuracy.

Critical weaknesses

1. Limited Methodological Novelty. The proposed framework primarily integrates existing techniques rather than introducing new methodological contributions. Dynamic graph modeling for capturing inter-sensor dependencies and structural anomalies has been extensively studied in prior work, including MADGA and SARAD. The main distinctions of TOAD lie in its latent-distribution-based anomaly scoring and score-aligned sensor attribution. However, both rely largely on established techniques, and their methodological novelty appears incremental.

2. Insufficient Experimental Validation of the Proposed Innovations. The experiments do not adequately establish the advantages of TOAD's key design choices over closely related methods. In particular, MADGA, which also employs dynamic graphs and optimal transport for anomaly detection, is notably absent from the baseline comparisons. Although the paper provides component-level ablations, it lacks direct comparisons between its proposed scoring mechanism and existing graph-level OT approaches, as well as sufficient ablations isolating the contributions of individual scoring signals and the proposed sensor-level attribution mechanism.

3. Methodological Inconsistencies and Lack of Clarity. The presentation contains inconsistencies that make the proposed framework difficult to understand and reproduce. Notably, Figure 1 illustrates reconstruction, temporal consistency, and optimal transport losses during training, whereas Eq. (6) defines a substantially different objective involving reconstruction, spectral regularization, contrastive learning, and independence losses. Furthermore, although the paper emphasizes counterfactual repair and graph-aware propagation as part of its attribution mechanism, their roles in the actual attribution computation are not clearly specified.

Questions for authors

See the weakness.

Could the authors clarify whether Figure 1 was generated using AI tools? The figure appears to contain inconsistencies with the methodology described in the paper.

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

This paper proposes TOAD, a framework for explainable multivariate time-series anomaly detection and root-cause localization. It constructs dynamic sensor graphs to model inter-sensor dependencies and combines sliced Wasserstein distance, reconstruction error, and temporal-transition error to detect anomalies. For root-cause localization, it uses gradient-based attribution and counterfactual repair to identify sensors contributing to anomalous scores. Experiments on five public datasets evaluate both anomaly detection performance and sensor-level localization.

Critical strengths

1.The paper addresses an important limitation of multivariate time-series anomaly detection: conventional reconstruction- and forecasting-based methods may overlook anomalies arising from changes in cross-sensor dependencies. TOAD combines dynamic sensor graphs with optimal-transport-based scoring and score-aligned sensor attribution to support both anomaly detection and root-cause localization. 2.The evaluation covers five public datasets, ten baselines, and multiple metrics. TOAD achieves the best threshold-independent detection performance on four of the five datasets. On SMD, it improves root-cause localization over the strongest reported baseline from 64.0% to 72.8% in H@100 and from 78.2% to 86.4% in H@150. The counterfactual repair experiments also provide an additional evaluation of explanation faithfulness.

Critical weaknesses

1.Insufficient experimental validation and novelty. Section 2 discusses MADGA and SARAD, but neither is evaluated as a direct baseline for detection or root-cause localization (Tables 1 and 4). Appendix B, Table 8 does not isolate the reconstruction and temporal-transition branches. Section 5 and Table 5 lack clear definitions of AOPC and (\Delta S@k) and justification for cross-model score comparisons. Reproducibility is also limited by underspecified graph construction (Section 3) and missing hyperparameters, including (\lambda_3) and (\epsilon_{\mathrm{adv}}) (Appendix B). Fig. 1 is cluttered, and the legend in Appendix Fig. 3 overlaps with data. 2.Potential gaps in theoretical guarantees and attribution. Section 3 recomputes the graph for each window, whereas Lemma 1 (Section 4) assumes a fixed normalized graph operator (\hat{A}_j). Theorem 1 relies on McDiarmid's inequality despite heavily overlapping windows with stride (S=1). Theorem 2 bounds false attribution to non-root-cause sensors but does not guarantee detection of true root causes. Moreover, the reconstruction attribution in Section 3, Eq. (5), may omit chain-rule terms through the input-dependent decoder. The final RCA ranking mechanism also remains unclear regarding gradient attribution, counterfactual score reduction, and graph propagation.

Questions for authors

1.Could the authors directly compare TOAD with MADGA and SARAD for anomaly detection and root-cause localization, and provide separate ablations for the reconstruction and temporal-transition branches? 2.Could the authors clearly define AOPC and (\Delta S@k), and explain how scores are normalized for fair cross-model comparisons? 3.Lemma 1 assumes a fixed normalized graph operator (\hat{A}j), whereas Section 3.1 constructs a dynamic graph for each input window. How do the theoretical guarantees apply to this input-dependent graph construction? 4.Theorem 1 uses McDiarmid's inequality despite overlapping windows at stride (S=1). Theorem 2 bounds false attributions but does not guarantee correct root-cause identification. Could the authors justify these theoretical guarantees under the actual setting? 5.Could the authors clarify the final RCA computation, including the reconstruction gradient and the use of counterfactual scores or graph propagation, and report missing hyperparameters such as (\lambda_3) and (\epsilon{\mathrm{adv}})?

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

Related papers

Discussion (0)

Sign in to comment.