PAI: Preserving Amplitude Information in Representation-Based Time-Series Anomaly Detection

2026-06-08 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

PAI, a new anomaly scoring scheme, addresses a critical limitation in representation-based time-series anomaly detection: their learned embeddings often lose amplitude information, degrading performance on amplitude-related anomalies. PAI introduces a diagnostic module that compares cosine and Euclidean scoring to assess amplitude capture, alongside a final score augmentation function. This function computes point-wise median and MAD deviation scores, plus a local mean-shift score, which are then fused with the base representation score. Evaluated on the TSB-AD-U-Eva and TAB UV datasets, PAI improved all four tested representation-based methods across every metric, yielding average VUS-PR gains of 98.4% and 36.8%. Specifically, PaAno + PAI achieved the best performance, surpassing the previous best method by 15%.

Key takeaway

For Machine Learning Engineers developing time-series anomaly detection systems, if your current representation-based methods struggle with amplitude-related anomalies, consider integrating PAI. This scheme explicitly preserves amplitude information, demonstrably improving VUS-PR gains by up to 98.4% on datasets like TSB-AD-U-Eva. You should evaluate PAI's impact on your specific datasets, especially when amplitude variations are critical for anomaly identification.

Key insights

Representation-based time-series anomaly detection benefits significantly from explicitly preserving amplitude information, which PAI achieves through a novel scoring scheme.

Principles

• Amplitude information is crucial for time-series anomaly detection.
• Representation-based methods often lose amplitude data.
• Fusing multiple deviation scores enhances anomaly detection.

Method

PAI's diagnostic module compares cosine and Euclidean scoring. Its augmentation function computes point-wise median/MAD deviation and local mean-shift scores, fusing them with the representation score.

In practice

• Integrate PAI with existing representation-based methods.
• Evaluate amplitude-related anomaly detection performance.
• Utilize the provided PAI code for implementation.

Topics

• Time-Series Anomaly Detection
• Representation Learning
• Amplitude Preservation
• PAI Scoring Scheme
• Anomaly Scoring
• PaAno

Code references

• pantheon5100/PAI

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

Related on AIssential

• Infinite-dimensional Mahalanobis Distance with Applications to Kernelized Novelty Detection — The Mahalanobis distance is extended to infinite-dimensional spaces via a variance norm for enhanced novelty detection.
• Self-Supervised Learning via Flow-Guided Neural Operator on Time-Series Data — FGNO enhances self-supervised learning for time-series data by dynamically varying corruption levels and using clean inputs for inference.
• Masked Diffusion Modeling for Anomaly Detection — MaskDiff-AD uses masked diffusion models to detect anomalies in discrete data by scoring reconstruction difficulty without reverse-time sampling.
• LATERN: Test-Time Context-Aware Explainable Video Anomaly Detection — LATERN enhances VLM-based video anomaly detection by integrating temporal context for coherent explanations.
• Starting Off on the Wrong Foot: Pitfalls in Data Preparation — Statistically informed data preparation significantly improves model robustness and efficiency for imbalanced, real-world datasets.
• ChronosAD: Leveraging Time Series Foundation Models for Accurate Anomaly Detection — ChronosAD leverages time series foundation models and a temporal block for robust, generalized anomaly detection with minimal tuning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.