Identifying bias in network clustering quality metrics.

Area of Science:

• Network science
• Data analysis
• Algorithm evaluation

Background:

• Network clustering quality metrics assess community structures.
• Existing metrics may exhibit biases related to internal and external connectivity.
• Evaluating these biases is crucial for accurate network analysis.

Purpose of the Study:

• To investigate potential biases in popular network clustering quality metrics.
• To develop a robust method for generating networks with controlled community structures and degree distributions.
• To introduce and evaluate a new quality metric, the density ratio.

Main Methods:

• Utilized stochastic and preferential attachment block models for network generation.
• Incorporated preset community structures, Poisson, and scale-free degree distributions.
• Generated multi-level structures to test metric performance across varying cluster numbers and strengths.

Main Results:

• Most evaluated metrics showed a bias towards favoring partitions with fewer, larger clusters.
• This bias persisted even when internal and external connectivity were comparable.
• The density ratio metric demonstrated reduced bias compared to other metrics.

Conclusions:

• Popular network clustering metrics often exhibit inherent biases.
• Modularity and the proposed density ratio metric appear less susceptible to these biases.
• Careful selection of quality metrics is essential for reliable community detection in networks.