Population recovery capabilities of 35 cluster analysis methods

Area of Science:

• Statistics
• Bioinformatics
• Computational Biology

Background:

• Cluster analysis is crucial for identifying population structures in complex datasets.
• Evaluating the performance of different clustering algorithms is essential for accurate data interpretation.
• Artificial data generation allows for controlled assessment of population recovery capabilities.

Purpose of the Study:

• To comparatively evaluate the population recovery capabilities of 35 distinct cluster analysis methods.
• To determine the optimal combinations of similarity measures and agglomeration rules for population recovery.
• To assess method performance under varying degrees of population overlap and profile differences.

Main Methods:

• Utilized artificial data representing duplicate mixture samples from 4 latent populations.
• Varied latent population mean profiles (elevation and pattern) and sampling variances (two overlap levels).
• Employed 5 profile similarity measures and 7 agglomeration rules to define 35 cluster analysis methods.

Main Results:

• Complete linkage and Ward's minimum variance methods demonstrated superior performance.
• Euclidian and city block interprofile distance measures were most effective with these top methods.
• Single linkage, median, and centroid methods showed significantly inferior results in population recovery.

Conclusions:

• Complete linkage and Ward's minimum variance, with specific distance measures, are recommended for robust population recovery.
• Method selection significantly impacts the accuracy of clustering individuals into their true population memberships.
• The study provides guidance for choosing effective cluster analysis techniques in similar research scenarios.