acceptodds
Withdrawn from ICLR 2027

PVBoost: Posterior-Calibrated Variance Boosting for Time-Series Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Time-series anomaly detection (TSAD) is typically performed in an unsupervised manner and follows a two-step pipeline: scoring, which quantifies the abnormality of each time step, and thresholding, which converts continuous scores into actionable decisions. Prior work largely emphasizes scoring, leaving thresholding to ad hoc heuristics despite their inherent coupling. This raises a key question: given anomaly scores from a black-box model, how can we refine the scores and select a decision threshold jointly, in a principled, unsupervised manner? To address this question, we propose \model, an unsupervised, parameter-light, and model-agnostic framework that jointly refines black-box anomaly scores and selects a decision threshold through a posterior-variance feedback loop. Mixture modeling estimates anomaly posteriors, which gate variance-based boosting to emphasize informative uncertainty while suppressing noise. Refined scores update the posteriors, whose trajectory also guides threshold selection, unifying refinement and decision-making without labels or access to detector internals. Experiments with statistical and neural detectors on an established TSAD benchmark comprising time series across nine domains demonstrate gains in both scoring and detection. In the univariate and multivariate settings, respectively, \model improves average VUS-PR by and over the original scores and average F-score by and over widely used three-sigma thresholding of those scores. These findings support coupled refinement and thresholding as a practical approach to more accurate unsupervised TSAD. Our implementation is available at https://anonymous.4open.science/r/PVBoost.

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

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_he89OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper attempts to address the problems of inaccurate anomaly scores and the difficulty of threshold selection in unsupervised time-series anomaly detection. The authors propose PVBoost, which refines anomaly scores by combining estimated anomaly probabilities with variations in predictions across iterations and automatically selects a detection threshold to improve anomaly detection accuracy.

Critical strengths
  1. Proposes a unified framework for anomaly score refinement and threshold selection, rather than treating the two steps independently.
  2. Uses anomaly probabilities to identify meaningful score variations, reducing the influence of noise in normal regions.
  3. Leverages the temporal continuity of anomalies to help automatically determine detection thresholds, reducing the need for manual threshold setting.
Critical weaknesses
  1. When the true anomaly ratio is below 2%, its performance is poor, even worse than the simple 3σ method.
  2. The paper claims that scoring and threshold selection can reinforce each other, but in reality, the threshold does not improve scoring in reverse.
  3. The improvement of score refinement is unstable, and some cases actually see performance deterioration.
  4. The method relies on the continuous occurrence of anomalies, which can easily cause misjudgments on scattered anomalies and sustained normal fluctuations.
  5. About 77% of the detection performance gain comes from threshold adjustment, and it remains uncertain whether the complex scoring optimization process is necessary.
Questions for authors
  1. Why is PVBoost not as effective as the simple 3σ rule when the true anomaly ratio is less than 2%? How can this problem be addressed?
  2. Since threshold selection does not feed back into score refinement, why is it called joint optimization? Can you provide direct evidence?
  3. Why do only 47% of detector–time-series pairs show improvement after score optimization? Can you explain the reason for the performance degradation?
  4. Can PVBoost maintain good detection performance when anomalies occur sporadically or when normal data exhibits continuous fluctuations?
  5. Since about 77% of the F-score improvement over the 3σ baseline comes from threshold adjustment, is complex iterative score optimization really necessary? Can you demonstrate that its additional benefits justify the computational cost?
Reviewer_p7H5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a method named PVBoost iteratively refine time series anomaly scores produced by black box models by combining the posterior of anomaly presence and the variance of the weak learners learning residuals. This is a shallow strategy fully unsupervised claiming to bridge score refinement and threshold selection as the threshold depends on the posterior trajectory. The authors claim that their method increase up to 11.4% VUS-PR and 107.3% F-score on respectively the original score and common three-sigma thresholding methods on the TSB-AD-U benchmark. In order to detail the impact of the components of PVBoost an ablation study is lead on boosting, gating with posterior, weighting with variance and mixture flexibility. Futhermore, the sensitivity against the learning rate and the momentum of EMA shows that standard choices are robust. Finally the case study tries to illustrates on examples the impact of the method.

Critical strengths
  1. The ablation study of section 4.2.2 appears to be rigorously lead and facilitate the understanding of the strength and weaknesses of the method.

  2. The methodological choices are convincingly motivated which makes the reading easier.

Critical weaknesses
  1. First, the lisibility of the detailed results is difficult. For instance in Table1 there are more than 300 scores but only 4 are emphasized in bold. I believe progress in this direction both in the main text and the appendix would be very beneficial.

  2. Scores refinement performances should be more nuanced in my opinion : AnomalyTrans seems to be the poor performing detector that PVBoost refines the best and the solely effect of this detector is only removed in Table16 in detailed results wherwheas the effect of the performance of this one detector is hidden in the main text to the best of my knowledge. Furthermore, Table16 shows that among detectors scoring above 0.3 AUC-PR, PCA and KMeansAD significatively decrease in performance while LSTMAD and DeepAnT are stable and the main text defends "Strong source scores generally remain stable" (line 393-394), there is here a margin of improvement in accuratly reflecting the results in the main text in my opinion. Finally, Table16 reports results in AUC-PR while the main text exclusively relies on VUS-PR.

  3. According to my understanding of the paper, the score refinement on its own plays a small role as displayed in Table16 and the important results of Table 14 in Appendix illustrate the moderate interaction between posterior and the thresholding (9% of the times series in TSB-AD-U and 17% in TSB-AD-M), the rest beeing the fragmentation test and the clipping of the posterior mass. This does not really defend the explanation of the paper that "posterior trajectory yields prior-free thresholds." (in conclusion).

Questions for authors

Q1. The Case Study is clear in the text but the graphs in Figure4 are hard to read to be convinced : it is not straightforward whether the yellow lines hides the blue or the green and this changes the conclusions. Could you adjust the plots in consequence ?

Q2. Could you adress the lisibility of the tables I mentioned earlier ?

Q3. Could you either argument to contest my understanding of the paper concerning the low interaction between the posterior trajectory and the threshold ? Or if the previous understanding is correct, could you nuance the main text claims on the threshold beeing induced by the posterior trajectory ?

Reviewer_XmpaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PVBoost, a parameter-light, model-agnostic framework built on a posterior-variance feedback loop. The raised question how jointly refining the scores and selecting a decision threshold is interesting, since thresholding also plays a critical role in anomaly detection. Experiments on TSB-AD with 15 detectors show the superity.

Critical strengths
  1. The proposed problem is important, since thresholding also plays a critical role in anomaly detection.
  2. this paper is well-written and has a clear structure.
Critical weaknesses
  1. The proposed solution is a bit complicated in terms of its process and it will lead to excessive time costs.
  2. When anomaly scores fail to reflect abnormal situations (in fact, most detectors have the problem of being unable to detect certain anomalies), this process will instead amplify false positives. This raises a critical question: would it be better to perform refinement directly on the original series?
  3. This paper relies on GMM, and a detailed comparison with GMM directly is necessary.
  4. The base detectors in this paper are too old, and thus cannot fully reflect the generalization of the methods.
  5. The overall design of the proposed method appears to be empirical, lacking theoretical guarantees. It is still unclear which part should be attributed to the improvement in the results.
Questions for authors

Please see the weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.