acceptodds
Under review as a conference paper at ICLR 2027

Screen Broadly, Diagnose in Context: Hierarchical Time Series Anomaly Detection

Abstract

Time series anomaly detection is essential to the reliability of cloud services, Internet of Things systems, and communication networks. Although many existing methods are sensitive to changes in data, statistical deviation alone does not establish whether a change reflects a genuine operational anomaly or benign variation. This gap between numerical sensitivity and operational meaning remains a key obstacle to industrial deployment. Large language models can incorporate operational semantics, but directly interpreting large volumes of numerical time-series data remains challenging. We introduce VeraTS, a hierarchical framework that addresses this gap through two complementary stages. The first, Time Series Perception for Signal Detection, screens broadly for candidate deviations while preserving event coverage on labeled development data. The second, Expert System for Semantic Judgment, translates operator policies into explicit rules for harmful direction and persistence, then deterministically verifies candidate deviations against these rules. We evaluate VeraTS with supplied policy programs on operational KPI series from backbone-link monitoring data. The frozen screen overlaps all evaluation events. VeraTS achieves state-of-the-art average anomaly detection performance on backbone links and the best results on nearly all evaluated datasets, with experiments validating its robustness and practical deployability.

Primary area:
learning on time series and dynamical systems
Keywords:
AIOps, Time Series Anomaly Detection

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_3k9DOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a two-stage anomaly detection pipeline for operational KPI data. The first stage combines nine direction-free deviation features into one score and picks the threshold that keeps as few observations as possible while missing no labeled development event. The second stage applies a rule that encodes the harmful direction, a rolling or seasonal reference, a robust threshold, a persistence requirement, and gap closing. An LLM can translate operator policies into such rules. On real backbone-link KPI data with 66 evaluation events, VeraTS ranks first in 10 of 12 category-metric columns against 34 baselines.

Critical strengths

The central point is real and practically important: the same waveform can be an anomaly under one policy and benign under another. The method is deterministic, auditable, and cheap. The ablations of direction and persistence (Fig. 5) and the parameter-sensitivity study are informative, and the baseline set is broad.

Critical weaknesses
  1. The comparison is not on equal terms. VeraTS receives a supplied policy that specifies the harmful direction, reference, persistence, and closing, and its first-stage threshold is selected on labeled development data. The baselines are generic detectors that never see the policy. A stronger controlled comparison would use each competitive baseline's scores as the first-stage candidate mask and then apply the same second-stage policy rule. That would separate the value of VeraTS's screening features from the value of the policy-conditioned diagnosis.

  2. The value of the first stage is not shown. No result reports the second stage without the gate, so it is unclear whether screening helps accuracy. The gate keeps 32.8% of observations, and Sec. 3.3 states that the current implementation still computes reference features over the full trace, so the cost argument in Eq. 10 is not realized in the evaluated system.

  3. The Event-F1* comparison is retrospective. For each method and category, the best percentile between 90 and 99 is chosen using evaluation labels (Sec. 4), and Fig. 4 marks VeraTS's "label-selected maxima". The F1 ablations in Fig. 5 are more permissive still, using unrestricted per-series maxima. This does not affect the VUS results, but it makes the four Event-F1* wins oracle operating points rather than deployable estimates. Event-F1* is also a point-adjusted metric, which is known to overestimate detection quality (Kim et al., AAAI 2022). Because width-48 closing can bridge predictions across long gaps, its interaction with the point-adjusted Event-F1* metric may further increase event-level scores.

  4. The abstract and introduction present LLM translation of operator policies as a stated component and contribution of the method, but the experiments use "supplied rules". No result in the main text measures whether the LLM produces correct rules.

  5. All real-data results come from a single dataset with 66 evaluation events, with no evaluation on a standard public benchmark. The main text does not say how many distinct incidents or services these events represent, and no uncertainty estimates are reported, so it is hard to judge how general the findings are.

Kim, S., Choi, K., Choi, H.-S., Lee, B., and Yoon, S. (2022). Towards a Rigorous Evaluation of Time-Series Anomaly Detection. In Proceedings of the AAAI Conference on Artificial Intelligence, 36(7), 7194–7201.

Questions for authors
  1. How do competitive baselines perform when their scores are used as the candidate mask for the same second-stage rule?
  2. What are the event-F1 results with thresholds fixed on development data rather than chosen on evaluation labels?
  3. How does the second stage perform without the first-stage gate, in accuracy and in runtime?
  4. How were the rule parameters selected, and on which data?
  5. How accurately does the LLM translate operator policies into rules, compared with a deterministic parser or hand-written rules?
  6. Can the screening stage be evaluated on a public anomaly detection benchmark?
Reviewer_wEpPOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes VeraTS, a hierarchical framework for time series anomaly detection with statistical anomaly screening and policy-based semantic judgment. In the first stage, the framework combines nine statistical features to identify candidate anomalous regions. In the second stage, it applies predefined operational rules concerning anomaly direction, magnitude, and persistence to filter these candidates and generate alert intervals. This paper introduced an optional LLM-based rule compiler to translate natural-language operational policies into structured executable rules. Experiments on real backbone-link monitoring data demonstrate strong performance compared with existing baselines.

Critical strengths
  1. The proposed method achieves competitive results across multiple evaluation metrics and outperforms a wide range of statistical, deep learning, foundation model, and LLM-based baselines.

  2. Incorporating domain-specific operational requirements into time series anomaly detection is a practically important direction for real-world applications.

Critical weaknesses
  1. Clarity and readability. Several parts of this paper are unnecessarily complicated and difficult to follow. For example, the two proposed modules are named Time Series Perception for Signal Detection and Expert System for Semantic Judgment. However, their actual operations are relatively straightforward: the first combines nine conventional statistical features to generate candidate anomaly masks, while the second applies predefined rules to evaluate anomaly direction, magnitude, and persistence. More direct terminology would make the methodology considerably easier to understand. I highly encourage the authors to carefully revise the paper, simplify unnecessary terminology, and explain the technical components more directly.

  2. Limited novelty. The proposed framework mainly combines existing statistical features, maximum-based score aggregation, thresholding, and rule-based temporal filtering. Although the overall combination is practically motivated, the individual components are relatively conventional. In particular, the authors should better justify the necessity of the hierarchical architecture and clarify its advantages over simpler approaches that directly apply operational rules to conventional anomaly detection methods.

  3. Limited evaluation on established public benchmarks. The main experiments rely on a real-world backbone-link monitoring dataset for which public access and reproducibility are not clearly established. I strongly encourage the authors to include experiments on at least two additional publicly available time series anomaly detection benchmarks. Such experiments would provide stronger evidence that the proposed approach generalizes beyond the specific operational setting considered in this paper.

Questions for authors
  1. How does the proposed system learn or infer operational semantics from data? My understanding is that the semantic judgment component primarily executes predefined rules rather than learning operational semantics directly from time series observations. Is this interpretation correct?

  2. Do the main experiments rely on manually supplied policy rules rather than LLM-generated rules? If so, what is the specific novelty and technical contribution of the semantic judgment component? How does it differ from conventional rule-based anomaly filtering?

  3. Could the authors provide more details about the dataset? In particular, I would appreciate a clearer description of its original source, collection process, construction procedure, and training/validation/test splits. Please also clarify whether the dataset will be publicly released and, if not, how independent researchers can reproduce the reported results.

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

The authors propose a two-stage framework for time series anomaly detection (TSAD). They propose to use a high recall first stage to and accommodate the context information to detect the anomalies.

Critical strengths
  1. The paper approaches an important problem that time series anomaly depends on context and can not be detected using only numerical patterns.
Critical weaknesses
  1. The stagewise ablations of the two-stage implementation are not fully clear from the paper. Table 3 shows stage 1-only Strict Event F1 of 0.070, which improves to 0.679 in the two-stage. Hence, it is unclear what gains stage 1 brings in. Stage 2 only ablations are not reported, making this difficult to assess.
  2. A broad set of prior works uses the two-stage principle of detecting candidate anomalies broadly first and then refining them. For example, Two-Stage Deep Anomaly Detection with Heterogeneous Time Series Data (Jeong et al. 2022), Hierarchical coarse-to-fine anomaly detection for high-dimensional time series (Lu et al. 2026), Argos: Agentic Time-Series Anomaly Detection with Autonomous Rule Generation via Large Language Models (Gu et al. 2025).
  3. In TSAD, label noise and inaccurate data are concerning [1], and hence, the cleaned-up corpus of TSB-AD [2] has been proposed for accurate evaluation. The paper does provide a TSB-AD evaluation in Appendix I, but it is performed on only 26 series, whereas there are 1700 time series.
  4. The paper has confusing, non-standard, unexplained terminologies like "Expert System for Semantic Judgment", "Sieve versus screening versus candidate gate", etc. The paper contains unnecessary jargon, making it difficult to follow.
  5. The paper has 168 pages, including the appendix, which did not allow me to read most of it, especially when we reviewers are not obliged to read such a long appendix. I encourage the authors to take time to present the paper concisely.

[1] Wu, Renjie, and Eamonn J. Keogh. "Current time series anomaly detection benchmarks are flawed and are creating the illusion of progress.

[2] Liu, Qinghua, and John Paparrizos. "The elephant in the room: Towards a reliable time-series anomaly detection benchmark." Advances in Neural Information Processing Systems 37 (2024): 108231-108261.

Questions for authors

Refer to the weaknesses.

Reviewer_kYs9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

VeraTS detects operational time-series anomalies in two stages. A direction-agnostic screen selects candidate deviations, and a deterministic verifier applies policies specifying harmful direction, reference baselines, and temporal persistence. The paper also describes an LLM interface for compiling natural-language policies into checked rules, although the main experiments use supplied rules. Experiments on backbone-link KPI data compare VeraTS with statistical, neural, and LLM-based detectors and examine how its components affect detection.

Critical strengths
  1. The paper explains why statistical deviation alone does not determine whether an operational alert is appropriate. A similar change in a signal can require different responses under different service policies. Making those policies explicit gives the detector information needed to interpret the signal and makes its decisions easier to inspect.
  2. The verifier produces reproducible decisions through explicit rules. The component analyses show how direction and persistence affect false alarms and how interval construction changes localization. These analyses help explain the reported performance beyond the aggregate comparison with other detectors.
Critical weaknesses

The main comparison does not establish how much improvement comes from the hierarchical design. VeraTS receives operator policies that most signal-only baselines do not receive, so their performance difference also reflects access to information about the target behavior. The matched-rule ablations are useful, but a simpler detector given the same policies would provide a more direct test of the proposed architecture. This matters for deciding whether to adopt the full pipeline or add policy conditioning to an existing detector.

The benefit of screening also needs to be isolated. Figure 4 reports measured CPU time, and the authors explicitly acknowledge that the current feature and verifier paths process the full trace. Retaining 32.8% of observations therefore does not establish a corresponding saving in computation. A comparison using the same verifier with and without candidate gating would show whether screening improves detection enough to justify its additional complexity and risk of excluding useful evidence. Preserving overlap with development events does not establish coverage of new failure types.

Generalization is difficult to assess because of the data dependencies reported in Appendix E.5. In particular, every Wave evaluation anomaly interval reuses the values at the corresponding position in the training sequence. Appendix N.11 selects thresholds without the held-out scenario labels, but the appendix also acknowledges that value reuse persists. This limits what the results establish about new incidents, even though the authors disclose the issue. The main experiments also evaluate supplied policies more directly than the benefit of using an LLM to compile them.

Questions for authors
  1. With identical policies and reference features, how do detection quality and measured runtime change when candidate gating is removed? A comparison with a simpler policy-conditioned detector would also help determine which gains require the hierarchical design.
  2. Given the value reuse documented in Appendix E.5, what evidence supports detection of previously unseen incidents? Please clarify whether any evaluation separates training and test data by independently verified incident or source identity, and which generalization claims the current scenario-based analysis can support.

Comments on the paper

1 comment
Area_Chair_zSWROpenReview ↗

Discussion Reminder

Dear Reviewers,

Thank you for your efforts so far. This is a friendly reminder to review the authors' rebuttal and respond to their comments as soon as possible, before the discussion deadline (November 18 AoE). Please also indicate whether your concerns have been addressed. Thank you for your time and cooperation!

Best regards,
AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.