# Calibrated One-class Classification for Unsupervised Time Series Anomaly Detection

**Authors:** Hongzuo Xu, Yijie Wang, Songlei Jian, Qing Liao, Yongjun Wang, Guansong Pang  
**arXiv:** [2207.12201](https://arxiv.org/abs/2207.12201)  
**Date:** 2022-07-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2207.12201v1/

## Key points

- Introduces a novel approach called calibrated one-class classification (COUTA) for unsupervised time series anomaly detection
- Existing methods can be deceived by unknown anomalies in training data, leading to biased and inaccurate normality boundaries
- COUTA proposes two calibrations for the one-class classifier: adaptive penalization of uncertain predictions and discrimination between normal samples and native anomaly examples
- Extensive experiments on six real-world datasets show that COUTA outperforms twelve state-of-the-art competitors with F1 score improvements ranging from 6% to 31%
- Scalability tests demonstrate COUTA's efficiency in terms of time series length and dimensionality
- Source code for COUTA is available on GitHub for further use by researchers or practitioners

## Summary

This paper introduces a novel approach called calibrated one-class classification (COUTA) for unsupervised time series anomaly detection. The goal of this task is to monitor and detect potential faults in various systems. Existing state-of-the-art methods focus on using advanced neural network structures and learning objectives to accurately learn normal patterns and behaviors in the data. However, these methods can be deceived by unknown anomalies present in the training data, leading to biased and inaccurate normality boundaries. To address this problem, COUTA proposes two calibrations for the one-class classifier. First, it adaptively penalizes uncertain predictions, reducing the impact of anomaly contamination while emphasizing confident predictions. Second, it discriminates between normal samples and native anomaly examples generated to simulate genuine abnormal behaviors based on the original data. These calibrations result in contamination-tolerant and anomaly-informed one-class learning, improving the modeling of normality. The authors conducted extensive experiments on six real-world datasets to evaluate the performance of COUTA compared to twelve state-of-the-art competitors. The results show that COUTA outperforms these competitors with F1 score improvements ranging from 6% to 31%. Additionally, scalability tests were performed to assess COUTA's efficiency in terms of time series length and dimensionality. Overall, this paper presents a novel approach that addresses the limitations of existing methods in unsupervised time series anomaly detection. The proposed COUTA method demonstrates improved performance and scalability compared to state-of-the-art competitors. Furthermore, source code for COUTA is available at a provided GitHub link for further use by researchers or practitioners interested in applying it for their own applications or research purposes.