Entropy-based dynamic graph embedding for anomaly detection on multiple climate time series.

Area of Science:

• Climate Science
• Data Science
• Machine Learning

Background:

• Abnormal climate events are characterized by extreme meteorological conditions.
• Current supervised learning methods fail to detect novel or untrained anomaly patterns.
• There is a need for advanced methods to accurately identify diverse climate anomalies.

Purpose of the Study:

• To propose a novel dynamic graph embedding model, EDynGE, for detecting abnormal climate events.
• To overcome the limitations of existing methods in identifying untrained anomaly patterns.
• To enhance the accuracy and scope of climate anomaly detection.

Main Methods:

• Constructing a dynamic graph by identifying correlations within climate time series data.
• Proposing EDynGE, a dynamic graph embedding model incorporating graph entropy.
• Quantifying graph information using graph entropy to build an effective embedding space.

Main Results:

• The EDynGE model demonstrated superior performance compared to baseline methods, achieving a 43.2% higher F1-score.
• Experiments on synthetic and real-world meteorological datasets validated the model's effectiveness.
• Analysis revealed a significant increase in abnormal climate event days over the past 30 years (304.5 days).

Conclusions:

• EDynGE effectively detects untrained abnormal climate event patterns, outperforming existing methods.
• The model's graph entropy-based approach provides a robust framework for anomaly detection in climate data.
• Findings highlight the increasing frequency of extreme climate events, underscoring the importance of advanced detection techniques.