Anomaly Detection in High-Dimensional Time Series Data with Scaled Bregman Divergence.

Area of Science:

• Data Science
• Machine Learning
• Statistics

Background:

• Anomaly detection identifies data points deviating from normal behavior, with broad applications.
• High-dimensional data presents challenges for existing anomaly detection algorithms.
• The unconstrained least-squares importance fitting (uLSIF) method struggles with unboundedness in certain high-dimensional scenarios.

Purpose of the Study:

• To propose a novel anomaly detection algorithm designed for high-dimensional data.
• To overcome the unboundedness problem encountered by existing methods like uLSIF.
• To enhance the accuracy and applicability of anomaly detection in complex datasets.

Main Methods:

• Developed a scaled Bregman divergence-based anomaly detection algorithm.
• Incorporated both least absolute deviation and least-squares loss for parameter learning.
• Evaluated the algorithm on synthetic and real-world high-dimensional time series datasets.

Main Results:

• The proposed algorithm effectively addresses the unboundedness problem.
• Demonstrated superior performance in detecting anomalies within high-dimensional time series data.
• Outperformed other density ratio estimation-based anomaly detection methods in comparative analyses.

Conclusions:

• The scaled Bregman divergence-based algorithm is a robust solution for anomaly detection in high-dimensional data.
• This method offers a significant improvement over existing techniques, particularly in challenging data environments.
• The algorithm's effectiveness on real-world datasets validates its practical utility.