Skip to content

ASTER: Latent Pseudo-Anomaly Generation for Unsupervised Time-Series Anomaly Detection

Home / Papers / ASTER: Latent Pseudo-Anomaly Generation for Unsupervised Time-Series Anomaly Detection

ASTER: Latent Pseudo-Anomaly Generation for Unsupervised Time-Series Anomaly Detection

Authors: Romain Hermary, Samet Hiçsönmez, Dan Pineau, Abd El Rahman Shabayek, Djamila Aouada Date: 2026-04-15 Paper ID: openalex:2604.13924

Summary

ASTER addresses the scarcity of labeled data in time-series anomaly detection by generating pseudo-anomalies directly within the latent space, bypassing the need for handcrafted anomaly injection. The framework utilizes a latent-space decoder combined with a pre-trained LLM to enrich temporal and contextual representations. This allows for the effective training of a Transformer-based classifier, leading to state-of-the-art performance across various benchmark datasets.

Key Contributions

  • Introduces ASTER, an unsupervised time-series anomaly detection framework that generates pseudo-anomalies in the latent space.
  • Leverages a pre-trained LLM to enhance temporal and contextual representations, improving upon standard reconstruction/forecasting methods.
  • Achieves state-of-the-art results on three major time-series anomaly detection benchmarks, surpassing existing embedding-based and reconstruction-based approaches.

Open Questions & Future Work

Key Concepts

  • latent-pseudo-anomaly-generation: A technique for unsupervised time-series anomaly detection that constructs pseudo-anomalies within the latent space, decoupling training from domain-specific hand-crafted rules.

Archivist Review

I have approved ‘Latent Pseudo-Anomaly Generation’ as a core technique, focusing on the mechanism rather than the specific framework name (ASTER). I also approved a refined version of the pseudo-anomaly diversity question, as this is a tangible bottleneck for unsupervised anomaly detection. Other candidates were either too generic or effectively renamed for long-term consistency in the vault.

Approved Concepts

  • Latent Pseudo-Anomaly Generation: It addresses the bottleneck of requiring handcrafted anomalies in unsupervised time-series anomaly detection by shifting the problem to the latent space, which is a highly reusable methodology for representation learning.

Approved Open Questions

  • Latent Pseudo-Anomaly Diversity Bottleneck: The robustness of unsupervised anomaly classifiers is intrinsically linked to the coverage of the synthetic anomaly distribution; improving this is central to model performance and generalization.

Rejected Candidates

  • [concept] ASTER (Latent Pseudo-Anomaly Generation) (aster-latent-pseudo-anomaly-generation) - other: Renamed to ‘Latent Pseudo-Anomaly Generation’ to focus on the underlying methodology rather than the specific framework name.
  • [open_question] Enhancing latent pseudo-anomaly diversity (enhancing-pseudo-anomaly-diversity-in-latent-space) - other: Renamed to follow a more concise, bottleneck-focused naming convention.
  • [open_question] LLM representational performance in TSAD (llm-representational-performance-in-time-series-ad) - generic: The question is too broad and generic regarding the application of LLMs to time-series, lacking a specific, actionable research hurdle distinct from general foundation model adaptation.

Metadata & Links

url
https://arxiv.org/abs/2604.13924
paper_id
2604.13924
paper_source
openalex
domain
time-series
tags
time-series-anomaly-detectionunsupervised-learningtransformerllm-for-time-series
architectures
datasets
concept_slugs
latent-pseudo-anomaly-generation
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-04-18T05:34:26Z