acceptodds
Withdrawn from ICLR 2027

Learning Time-Series Anomaly Generators from Synthetic Supervision

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Synthetic anomalies are widely used to alleviate the scarcity of labeled anomalies in time-series anomaly detection. Existing anomaly injection methods use predefined transformations to construct injected anomalies, which may limit data diversity and fail to sufficiently capture the temporal characteristics of the target data. This work proposes a generative framework for multivariate time-series anomaly synthesis based on normal-pattern pre-training and anomaly-oriented fine-tuning. The generator is first pre-trained exclusively on normal time-series data to capture target-specific temporal patterns. Then, multi-type injected anomalies are constructed to provide *synthetic supervision*, and a dimension-adaptive reconstruction objective is introduced during fine-tuning to adjust relative reconstruction penalties according to channel-wise statistics. After fine-tuning, latent perturbations broaden the sampling region of the learned latent distribution, enabling more diverse anomaly generation. Unlike conventional anomaly injection, the injected anomalies are not directly used as the final generated anomalies, but instead guide the generator in learning heterogeneous deviations from normal temporal patterns. Experiments on five multivariate time-series benchmarks and three downstream anomaly detectors show consistent improvements, with average AUPR gains of 13.28%, 11.28%, and 16.85% for CARLA, COUTA, and DACAD, respectively. Representation analysis further shows that the generated anomalies achieve stronger separability from normal patterns and greater diversity than direct anomaly injection, supporting their effectiveness for downstream training.

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

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_igQrOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a generative framework for time-series anomaly synthesis by learning from rule-based injected anomalies. Instead of directly using the injected anomalies as training samples for downstream anomaly detectors, the proposed method treats them as synthetic supervision to adapt a generative model, with the goal of producing more diverse and informative anomaly patterns. Specifically, the generator is first pre-trained on normal time series, then fine-tuned using multiple types of injected anomalies with a dimension-adaptive reconstruction objective, and finally used to generate new anomalous samples through latent perturbation.

Critical strengths
  1. The overall motivation is meaningful. Learning from rule-based anomaly generation mechanisms can be viewed as a way of incorporating useful inductive biases about abnormal patterns, rather than discarding such prior knowledge entirely. I find this direction potentially valuable, especially when real anomalous samples are scarce.

  2. The proposed framework is structurally simple and easy to understand. The method mainly consists of normal-pattern pre-training, anomaly-oriented fine-tuning, and latent perturbation. This relatively simple design makes the central idea of the paper clear and allows the contribution to be examined without being obscured by an overly complicated architecture.

Critical weaknesses
  1. The distinction from existing anomaly synthesis methods is not sufficiently convincing to me. Existing works such as GenIAS, CAAE, and CAROTS already explore ways of producing more informative or diverse synthetic anomalies, and there appears to be substantial overlap in the high-level motivation. The paper argues that directly using injected anomalies may limit anomaly diversity, and therefore proposes to use injected anomalies as supervision for a generator instead. However, I am not fully convinced that this constitutes a sufficiently fundamental distinction. In both cases, the injected anomalies still serve as training data from which a model learns a representation or distribution of abnormal patterns. A downstream anomaly detector trained directly on injected anomalies can also learn a distribution induced by these samples.

  2. Most evidence for the effectiveness of the generated anomalies is indirect and based on downstream performance. The main evidence comes from improved AUPR/F1 on downstream anomaly detectors and representation-space metrics such as separability and diversity. While these results are useful, I consider diversity in latent space to be only a weak surrogate for the actual quality of generated anomalies. Higher diversity does not necessarily imply that the generated samples are realistic, meaningful, or representative of plausible anomaly mechanisms. More direct evidence about the quality of the generated anomalies would strengthen the paper considerably.

  3. The evaluation provides limited morphological evidence for the generated anomalies. All evaluated datasets are multivariate, and the paper does not include a univariate evaluation where anomaly morphology can be inspected more directly. In addition, although the paper provides t-SNE visualizations in the representation space, it does not show qualitative comparisons of the generated time-series waveforms themselves. For a paper centered on anomaly generation, visualizing normal samples, rule-based injected anomalies, and generated anomalies in the original signal space would provide more intuitive evidence that the proposed generator indeed produces meaningful and structurally diverse anomaly patterns.

Questions for authors

See weaknesses.

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

The paper proposes a TCN-VAE anomaly generator for multivariate time series. It pretrains on normal windows, fine-tunes on five types of injected anomalies with channel-adaptive reconstruction weights, then perturbs latent codes to generate training anomalies. Across five benchmarks and three downstream detectors, it reports improved AUPR over detector-native injection and three synthesis baselines (GenIAS, CAAE, and CAROTS), with component ablations and runtime results.

Critical strengths
  • The work evaluates three synthetic-data methods across three downstream detectors and five datasets, with component ablations and runtime costs. This gives useful evidence that learned synthetic data can improve detector training.
  • The normal pretraining followed by anomaly-guided fine-tuning is a plausible and practically relevant design, and the reported gains over the corresponding detector baselines are substantial on average.
Critical weaknesses
  • Limited methodological novelty over prior latent-space generation. GenIAS also learns a VAE from normal data and perturbs its latent space to synthesize anomalies. This paper adds anomaly-supervised fine-tuning and channel-adaptive reconstruction, but does not clearly isolate what new anomaly-generation capability these additions provide beyond downstream detector gains. The conceptual advance therefore appears incremental.
  • Weak direct validation of generated-anomaly quality. The claim that the generator learns structured, heterogeneous anomalies is supported mainly by separability and pairwise-diversity distances from a normal-trained encoder on two MSL entities (Table 3). Both quantities can increase with perturbation magnitude and do not establish that generated windows resemble meaningful anomalies. The paper shows neither representative generated windows nor comparisons with held-out real anomalies.
  • Insufficient statistical and protocol detail for the performance claims. Table 2 reports substantial variability for MSL/SMAP/SMD but does not clarify whether it reflects variation across entities or random seeds, and gives no uncertainty for aggregate gains. The number of generated windows, injected-operator settings, and whether synthetic sample counts are matched across methods are also unclear. These details are needed to assess the stability and fairness of the reported improvements.
Questions for authors
  • How are generated windows assigned anomaly labels, and what evidence shows that they are anomalous rather than perturbed but still normal reconstructions? Do you have results against held-out real anomalies or qualitative samples beyond the two MSL entities?
  • For Table 2, what are the number of independent runs and the source of the reported standard deviations? Were generated-sample counts and injection settings matched across all synthesis methods?
Reviewer_9ZsxOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

his paper introduces a generator-based TSAD framework that leverages a pseudo-anomaly generator to generate pseudo anomalies for training. The high-level idea is quite similar to that of existing generator-based TSAD methods, as the authors themselves have discussed and compared. Specifically, similar to GenIAS, a TCN-VAE is used to model normal windows and is then fine-tuned using injected anomalies. The generator is subsequently used to improve the training of the detector. Experiments are conducted on five datasets to validate the performance of the proposed method.

Critical strengths
  1. The paper's writtening general clear and easy to follow.
  2. On the selected datasets, the proposed method achieves substantial improvements in TSAD performance when integrated with three different detection models.
Critical weaknesses
  1. The paper itself relies on the rule-based anomaly injection methods it criticises. Specifically, rule-based anomaly injection is used to fine-tune the generator.
  2. The injection process does not have an anomaly-specific objective. Moreover, latent noise is applied to the entire window, resulting in window-level anomalies rather than anomalies confined to specific time steps or channels.
  3. I acknowledge that the proposed method differs from GenIAS in several design choices, but their high-level workflows are similar. I do not find these design choices well justified, nor do I see any clear advantages over GenIAS.
  4. Would the proposed strategies also improve performance on univariate TSAD?
  5. The performance improvements are fairly moderate, and the proposed method does not outperform GenIAS in many cases.
Questions for authors

Please see my weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.