Dynamic graph embedding for outlier detection on multiple meteorological time series.

Area of Science:

• Data Science
• Climate Science
• Graph Theory

Background:

• Current dynamic graph embedding methods overlook graph similarities, limiting outlier detection in complex datasets.
• Detecting abnormal climatic events from meteorological time series requires methods that consider both temporal evolution and event similarity.

Purpose of the Study:

• To propose DynGPE, a dynamic graph embedding model incorporating graph proximity for enhanced outlier detection.
• To effectively identify abnormal climatic events within meteorological time series data.

Main Methods:

• Representing climatic events as graphs with meteorological data as vertices and spurious relationships as edges.
• Utilizing graph proximity (distance between graphs) to cluster similar climatic events in an embedding space.
• Applying established outlier detection algorithms (Isolation Forest, Local Outlier Factor, Box Plot) to identify distant events.

Main Results:

• DynGPE demonstrated superior performance, achieving an average F-measure improvement of 44.3% over baseline methods.
• Isolation Forest emerged as the most effective and stable outlier detection method, outperforming Local Outlier Factor and Box Plot.
• Specific performance gains against other methods were noted: 15.4% over Local Outlier Factor and 78.9% over Box Plot.

Conclusions:

• DynGPE effectively addresses the limitations of existing methods by incorporating graph similarity for outlier detection.
• The proposed model enables robust identification of abnormal climatic events in meteorological time series.
• Isolation Forest is recommended for its high performance and stability in conjunction with DynGPE for climatic event outlier detection.