acceptodds
Under review as a conference paper at ICLR 2027

Graphs on STAGE for Interpretable Time Series Anomaly Detection

Abstract

Multivariate time series monitoring is critical for safeguarding complex systems, where anomalies typically stem from interactions among interdependent dimensions. However, current detection paradigms face an inherent trade-off: traditional graph methods offer interpretability, but lag in empirical performance, while modern deep learning models provide high expressiveness at the cost of speed and transparency. We present STAGE, a self-supervised framework combining representation learning with explicit relational structure. STAGE models the time series dimensions as vertices in a dynamic graph, pruning edges via a multi-head attention mechanism to isolate genuine cross-channel dependencies. For explainability, STAGE derives instantaneous, post-hoc attributions by occluding individual dimension nodes and observing the shift in anomaly scores, requiring zero retraining. On the largest benchmark, TSB-AD-M, Stage achieves a VUS-PR of 0.44, matching the best baseline (out of 30) and outperforming the runner-up by 12%. Crucially, it matches both the detection accuracy and diagnostic interpretability of state-of-the-art methods, which cannot provide real-time explanations.

Primary area:
learning on time series and dynamical systems
Keywords:
Anomaly detection, Multivariate time series, Graph transformer, Explainable Anomaly detection

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Z9gHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

STAGE is a framework for multivariate time series anomaly detection with interactive channel attribution. Each temporal window becomes a graph whose nodes represent channels. Multi-head attention and pruning construct the graph, graph convolutions model dependencies, and a permutation-invariant Deep Sets readout produces an embedding trained with a Deep SVDD objective. Channel attributions are computed from the change in anomaly score after deleting each node, with the perturbed graphs evaluated in a batch without retraining. On the 180-series TSB-AD-M evaluation subset, STAGE obtains mean VUS-PR of 0.44, matching StreamVAE among 30 baselines. Experiments on CoMuT and SMD additionally evaluate attribution quality, explanation latency, training time, and an interactive visualization tool.

Critical strengths
  1. A practical integration of detection and interactive attribution. The graph representation permits channel deletion and immediate recomputation without retraining. Combining this capability with a permutation-invariant readout and an interactive interface offers a useful way for users to inspect multivariate alerts. The contribution has practical value even without a new detection objective.

  2. Competitive detection across a broad benchmark. The reported VUS-PR of 0.44 matches the strongest baseline on the 180-series evaluation subset. The paper separates tuning and evaluation subsets, averages its computed metrics over five runs, and includes rank-based statistical comparisons. These experiments provide substantially stronger support for the detector than evaluation on a few selected datasets would provide (Section 5 and Appendix B).

Critical weaknesses
  1. The central explanation-speed comparison does not isolate the benefit of the proposed architecture. Section 5.3 compares STAGE's batched node-deletion scores with SHAP explanations for competing detectors, reporting 0.03/0.05 ms versus 235/592 ms for StreamVAE on CoMuT/SMD. This compares different attribution procedures as well as different detectors. Reconstruction/prediction residuals can already provide inexpensive channel scores, as the related-work discussion acknowledges. Fixed-dimensional detectors can also be evaluated with batched input masking, although masking and deleting a graph node have different semantics. Without these controls, the experiments establish a large advantage over the selected SHAP implementation, but do not establish the broader claim that competing methods cannot offer real-time explanations. This distinction matters because explanation efficiency is the paper's principal advance over a detector with matching VUS-PR.

  2. The diagnostic interpretation needs stronger attribution controls. Deleting a node changes the graph presented to the model and may change its aggregation statistics; permutation invariance does not establish that scores remain calibrated under such changes in graph size. The score difference is well defined as a model sensitivity, but the paper does not isolate the effect of removing anomalous channel content from generic effects of deleting a node. This matters for interpreting the scores as evidence for root-cause analysis. Attribution evaluation on CoMuT and SMD is relevant, but Appendix B.2 computes IPS over anomalies detected by each model and weights them by detected points. Different detectors can therefore be evaluated on different event subsets. A comparison on common detected events, together with a control for generic score changes caused by deletion, would provide more convincing support for the diagnostic claims. This gap affects the interpretation of the explanations, even if the detection results remain valid.

Questions for authors
  1. Explanation comparison. On the existing CoMuT/SMD experiments, how does STAGE compare with native per-channel residual scores and batched channel masking for the strongest applicable baselines? Please specify the masking reference, attribution quality, common hardware, and whether the quoted times are single-window latency or amortized throughput, including the timing/synchronization procedure. A convincing advantage under these controls would materially strengthen the main contribution.

  2. Diagnostic validity. What evidence distinguishes the proposed deletion scores from sensitivity to changing graph size? Can the authors report attribution quality on common detected anomaly events, with detection coverage stated, and clarify which claims concern model attribution versus causal root-cause identification? Existing CoMuT/SMD results under these controls could resolve this concern without adding a new benchmark.

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

This paper addresses the tension between interpretability and performance in multivariate time series anomaly detection (MTSAD) and proposes a self-supervised framework, STAGE: it models each dimension of a multivariate series as a node in a dynamic graph and learns to characterize cross-dimensional dependencies via multi-head attention; it uses Deep SVDD as the one-class classification loss; for interpretability, it leverages the graph's native structure to perform post-hoc attribution via node-wise occlusion without retraining, and is accompanied by visualization tools. On the largest-scale benchmark TSB-AD-M, STAGE achieves a VUS-PR of 0.44, on par with the current SOTA and 12% higher than the third-place method, while also being faster in training.

Critical strengths
  1. The method is novel and insightful: the combination of "dynamic graph + attention pruning + DeepSet permutation invariance + Deep SVDD" is a structurally clean design.
  2. Native interpretability: the occlusion attribution leverages the graph structure and can be completed in a single batched forward pass
  3. The experiments are solid: the largest benchmark TSB-AD-M, 30 baselines.
  4. The efficiency advantage is significant: the speed are improved by 1–4, which has practical deployment value.
Critical weaknesses
  1. Evidence for interpretation quality is insufficient: IPS is comparable to StreamVAE, and it is not demonstrated that the interpretations are better; the interpretation comparison is limited to only two datasets, and the occlusion method (removing one node at a time) captures only univariate contributions, making it difficult to explain cross-dimensional joint anomalies.
  2. The fairness of the efficiency comparison is questionable: there is a lack of comparison with methods that are also "natively interpretable."
Questions for authors
  1. Occlusion attribution removes only one node at a time. How can it characterize cross-dimensional anomalies that only become apparent when multiple dimensions are jointly considered? Can you provide experiments or analysis on multi-node joint occlusion (or higher-order attribution)?
  2. The interpretation comparison is conducted only on SMD and CoMuT. Can you supplement IPS comparisons on the annotated subset of TSB-AD-M or on more datasets?
  3. On data with concept drift or non-stationarity, can the window-wise dynamic graph still perform stable pruning? Are there relevant experiments?
Reviewer_fgRwOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes STAGE, a framework for interpretable multivariate time-series anomaly detection. Each sliding window is represented as a graph whose nodes correspond to individual channels. Attention-based edge pruning, graph convolution, and permutation-invariant aggregation produce a representation that is scored using a Deep SVDD objective. Channel attributions are obtained by measuring changes in the anomaly score after removing individual nodes, without retraining the detector. The paper evaluates detection performance on TSB-AD-M, studies attribution quality and computational cost on CoMuT and SMD, and presents an interactive visualization for inspecting anomalous windows, graph structure, and channel contributions.

Critical strengths

1-The evaluation covers 180 multivariate time series in TSB-AD-M-Eval and compares STAGE against 30 baselines. Its reported mean VUS-PR of 0.44 matches that of StreamVAE, the strongest baseline by this aggregate measure. The accompanying statistical analysis provides further evidence that STAGE belongs to the leading group of methods. This breadth is a meaningful strength: the proposed framework offers an explanation mechanism and an interactive interface while retaining competitive detection performance across a substantial benchmark.

2-Representing channels as graph nodes makes node-removal attribution a natural operation. The same trained model can score the reduced graphs, and the perturbed graphs can be evaluated in a batch. The visualization further connects time-series observations, learned relationships, anomaly scores, and channel attributions, allowing users to inspect windows and examine the effects of removing nodes. Although the individual methodological components are largely established, their integration is useful. The reported results suggest an attractive trade-off between attribution quality and computational cost, with the interface adding practical value beyond aggregate detection metrics.

Critical weaknesses

1-Section 5.1 evaluates competing detectors using SHAP, whereas STAGE uses batched leave-one-node-out scoring. The substantial speedups therefore reflect differences in both the detector and the explanation algorithm. They establish an advantage over the evaluated detector–SHAP combinations, but leave open how STAGE compares with less expensive alternatives.

In particular, comparisons against simple channel masking or replacement, and native per-channel residual explanations where available, would better support the practical efficiency claim. The paper itself acknowledges that prediction-based methods can provide residual explanations. These alternatives need not have identical semantics to node deletion, but they are relevant to the stated trade-off among detection accuracy, explanation quality, and cost. Attribution quality should be reported alongside runtime, since Figure 4 shows that STAGE has lower IPS than some competing methods.

2-The score in Definition 2, (d_j=s(G)-s(G\setminus v_j)), measures how the learned anomaly score changes when a node and its incident edges are removed. This is a useful measure of model sensitivity, but it does not by itself identify which variable caused an anomaly in the underlying system. This distinction matters particularly when channels are correlated: a highly attributed channel may reflect a cause, a downstream effect, or a correlated signal used by the detector.

The discussion of “causal impact” in Section 2 and “root-cause analysis” in the conclusion would benefit from explicitly distinguishing interventions on model inputs from interventions on the monitored system. The CoMuT examples and IPS results provide evidence for anomalous-channel localization. Moreover, Appendix B.2 specifies that IPS is conditional on successfully detected anomalies, which limits its interpretation as an overall diagnostic measure. I view the visualization as a useful tool for inspection and hypothesis generation; stronger claims about causal identification require additional justification.

Questions for authors

1-Can the authors compare STAGE with simple channel-masking approaches under comparable evaluation budgets, and with native residual-based explanations on representative baselines? Reporting both IPS and runtime would help establish the practical trade-off more convincingly.

2-Do the reported explanation times of 0.03 ms and 0.05 ms per window represent individual-request latency or amortized throughput across batches? Please clarify whether perturbation generation and graph construction are included, specify the SHAP configuration and evaluation budget, and explain how runtime and memory scale with the number of channels, given that STAGE evaluates a batch of (D) perturbed graphs.

3-When a node is removed, are attention weights, pruning decisions, and the remaining-node representations recomputed? How should the resulting attribution be interpreted when several channels are strongly correlated, and what evidence supports extending that interpretation from contribution to the model’s score to the underlying system’s root cause?

Comments on the paper

1 comment

We sincerely thank the reviewers for their time and thoughtful feedback. We are pleased that all reviewers acknowledged several strengths of our work and awarded high scores for its presentation.

We have already uploaded a revised version of our PDF that includes additions and improvements suggested by the reviewers. Each addition is color-coded by the reviewer that suggested it, for ease of reading. Specifically:

  • edits made in response to reviewer fgRw are highlighted in BLUE
  • ORANGE for reviewer XvKB
  • RED for reviewer Z9gH.

We believe we have clarified all questions raised by the reviewers, but we remain available for further clarifications.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.