GraphTS: Graph-represented time series for subsequence anomaly detection.

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Subsequence anomaly detection in time series is crucial across many fields.
• Existing methods often require knowing anomaly length and count, and struggle with recurrent anomalies.
• Current approaches may fail to capture recurrent subsequence anomalies due to reliance on local information.

Purpose of the Study:

• To propose a novel graph-represented time series (GraphTS) method for subsequence anomaly discovery.
• To address limitations of existing methods, including the need for prior knowledge of anomaly characteristics.
• To effectively capture both recurrent and rare subsequence anomalies of arbitrary lengths.

Main Methods:

• Introduced a new time series graph representation model (GraphTS).
• Developed a 2D time series visualization (2Dviz) method to map 1D patterns into a 2D spatial-temporal space.
• Constructed a graph from the 2D representation to identify recurrent and rare subsequence anomalies.

Main Results:

• The GraphTS method successfully represents both recurrent and rare time series patterns.
• The 2Dviz technique enhances the recognition of subsequence anomalies.
• Experimental results show GraphTS outperforms state-of-the-art methods in accuracy and efficiency.

Conclusions:

• The proposed GraphTS method offers an effective approach for discovering single and recurrent subsequence anomalies.
• GraphTS overcomes limitations of existing methods by not requiring prior knowledge of anomaly length or quantity.
• The method demonstrates superior performance compared to current state-of-the-art techniques.