Assessing the robustness of cluster solutions obtained from sparse count matrices.

Area of Science:

• Psychology
• Data Science
• Network Analysis

Background:

• Psychological research increasingly uses community detection for clustering sparse count matrices (e.g., social networks, brain imaging).
• Existing methods lack robust approaches to evaluate the reliability of these cluster solutions in empirical data, leaving results susceptible to noise.

Purpose of the Study:

• To introduce two novel methods for evaluating the robustness of cluster solutions derived from community detection in psychological research.
• To present the R package, perturbR, which automates these robustness assessment techniques for psychological researchers.

Main Methods:

• Method 1: Compares original cluster assignments with those derived from randomly perturbing matrix edges to assess stability.
• Method 2: Utilizes Monte Carlo simulations to compare the modularity score of the original solution against random matrices with similar properties.
• The perturbR package implements these methods, facilitating their application in psychological studies.

Main Results:

• Demonstrated the utility of perturbR using benchmark simulated data.
• Applied the robustness assessment methods to publicly available empirical data from social networks and structural neuroimaging.

Conclusions:

• The introduced methods and perturbR package provide psychological researchers with essential tools to validate community detection cluster solutions.
• These approaches enable researchers to distinguish reliable findings from those potentially caused by random data fluctuations, enhancing scientific rigor.