RoCA: Robust Contrastive One-Class Time Series Anomaly Detection With Contaminated Data.

Area of Science:

• Machine Learning
• Data Science
• Signal Processing

Background:

• Time series anomaly detection (TSAD) is crucial due to increasing data volumes and label scarcity.
• Existing methods struggle with incomplete normality assumptions and data contamination, limiting robustness.

Purpose of the Study:

• To propose RoCA, a unified and robust framework for TSAD that addresses assumption incompleteness and data contamination.
• To develop a method that does not rely on clean training data.

Main Methods:

• RoCA employs a composite loss function with multinormality alignment, dynamic anomaly awareness, and variance terms.
• It leverages the insight that normal samples satisfy multiple assumptions, while anomalies violate at least one.
• The framework dynamically identifies and excludes anomalies during training to refine detection boundaries.

Main Results:

• RoCA consistently outperforms state-of-the-art methods on univariate and multivariate time series datasets.
• Achieved up to 7.3% improvement under real-world data contamination.
• Demonstrated the synergy between contrastive learning (CL) and one-class classification (OC) within the RoCA framework.

Conclusions:

• RoCA offers a robust and effective solution for TSAD, overcoming key limitations of existing approaches.
• The framework's ability to handle contaminated data and incomplete assumptions makes it suitable for real-world applications.
• RoCA provides a unified approach, enhancing both theoretical understanding and practical performance in anomaly detection.