Towards explainable community finding

Area of Science:

• Network Science
• Machine Learning
• Data Mining

Background:

• Community detection is crucial for network analysis, with many algorithms used in fields like public health.
• Existing methods often lack clear explanations for their community assignments.
• Understanding the reasoning behind community labels is essential for trust and application.

Purpose of the Study:

• To develop a model-agnostic methodology for post-hoc explanations of community detection algorithms.
• To identify informative network features that elucidate community structures.
• To provide insights into the commonalities and differences between various community detection approaches.

Main Methods:

• Proposing a novel methodology inspired by machine learning interpretability techniques.
• Applying the methodology to explain outputs from three well-established community detection algorithms.
• Identifying and analyzing a set of interpretable network features.

Main Results:

• The methodology successfully provides post-hoc explanations for community detection algorithms.
• Key network features contributing to community assignments were identified.
• Common and distinct explanatory features across different algorithms were reported.

Conclusions:

• The proposed methodology offers a generalizable approach to explain community detection outputs.
• Understanding the features driving community formation enhances the interpretability and applicability of network analysis.
• This work bridges the gap between complex algorithms and actionable insights in network science.