Concept Drift and Anomaly Detection in Graph Streams.

Area of Science:

• Graph theory and network analysis
• Stochastic processes and time series analysis
• Machine learning and data mining

Background:

• Graph representations are widely used across various domains.
• Detecting changes in the underlying process generating graph data is crucial for many applications.
• Existing methods often struggle with graphs of varying sizes and without vertex correspondence.

Purpose of the Study:

• To propose a general methodology for detecting changes in stationarity of stochastic processes generating attributed graphs.
• To handle graphs with a variable number of vertices and edges.
• To enable change detection without assuming vertex correspondence across time steps.

Main Methods:

• Embedding attributed graphs into a vector domain.
• Applying conventional multivariate change detection procedures in the vector space.
• Theoretical analysis to support the methodology's soundness.
• Implementation and evaluation on real-world datasets.

Main Results:

• The proposed methodology effectively detects changes in stationarity for graph-generating processes.
• The approach is robust to variations in graph size and vertex structure.
• Experimental results on biological molecules and drawings demonstrate superior performance compared to baseline methods.

Conclusions:

• The developed methodology provides a powerful and general tool for change detection in streaming attributed graph data.
• This approach enhances the analysis of dynamic graph data in various scientific and technical fields.
• The vector embedding technique offers a flexible framework for applying established change detection algorithms to graph streams.