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

2026-04-16 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, long

Summary

ASTER is a novel framework for unsupervised time-series anomaly detection (TSAD) that addresses challenges like rare anomalies and scarce labeled data by generating pseudo-anomalies directly in the latent space. It avoids manual anomaly injection and domain-specific expertise by using a latent-space decoder to produce tailored pseudo-anomalies, which then train a Transformer-based anomaly classifier. A pre-trained Large Language Model (LLM) enriches the temporal and contextual representations within this latent space. Evaluated on four benchmark datasets—PSM, PUMP, SWaT, and CATSv2—ASTER achieved state-of-the-art performance, significantly improving $\textrm{F}_{1}$-scores by over 0.3 points on PSM and SWaT datasets compared to previous LLM-based methods. The framework uses LoRA fine-tuning for the LLM and a Transformer-based classifier, demonstrating its effectiveness in modeling sequential data.

Key takeaway

For research scientists developing unsupervised time-series anomaly detection systems, you should investigate ASTER's approach of latent pseudo-anomaly generation and LLM integration. This method offers a robust alternative to traditional reconstruction or embedding-based techniques, potentially yielding superior performance on complex datasets. Consider adopting LoRA fine-tuning for pre-trained LLMs to enhance contextual understanding and improve classification accuracy in your models.

Key insights

ASTER generates latent pseudo-anomalies to train a Transformer-based classifier, leveraging LLMs for state-of-the-art unsupervised time-series anomaly detection.

Principles

• Generate pseudo-anomalies in latent space.
• Use LLMs for temporal and contextual enrichment.
• Employ Transformer-based classifiers for robustness.

Method

ASTER employs a latent-space decoder to generate pseudo-anomalies, which then train a Transformer-based classifier. A pre-trained LLM, fine-tuned with LoRA, enriches latent representations, enabling autonomous, robust anomaly detection without handcrafted injections.

In practice

• Apply LoRA for LLM fine-tuning in TSAD.
• Utilize VUS metrics for robust evaluation.
• Consider window size 4 for optimal performance.

Topics

• Time-Series Anomaly Detection
• Pseudo-Anomaly Generation
• Latent Space
• Large Language Models
• Transformer Architecture

Code references

• uniluxembourg/snt

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• ASTER: Latent Pseudo-Anomaly Generation for Unsupervised Time-Series Anomaly Detection — ASTER generates latent pseudo-anomalies to train a Transformer-based classifier, enhancing unsupervised time-series anomaly detection.
• Beyond Temperature: Hyperfitting as a Late-Stage Geometric Expansion — "Hyperfitting" improves LLM generation by dynamic rank reordering, distinct from temperature scaling, localized to the final transformer block.
• Efficient ASR Training with Conversations that Never Happened — LLM-generated synthetic conversations, synthesized via TTS, effectively augment scarce real data for ASR training.
• Understanding Key Features of Time Series Foundation Models from Epidemic Forecasting — Heterogeneous pretrained representations improve epidemic forecasting, especially with aligned pretraining domains.
• Trait-space Monitoring for Emergent Misalignment During Supervised Finetuning — Emergent misalignment in LLMs can be detected by monitoring representational drift in trait-space during supervised finetuning, offering a low-overhead alternative.
• MILM: Large Language Models for Multimodal Irregular Time Series with Informative Sampling — LLMs can effectively process multimodal irregular time series by encoding sampling patterns and values in XML.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.