Hierarchical Cluster Analysis Using Intraclass Correlations: A Mixture Model Study.

Area of Science:

• * Statistics and Data Mining
• * Machine Learning Algorithms
• * Cluster Analysis

Background:

• * Hierarchical clustering is a common method for data partitioning.
• * Evaluating algorithm performance on mixture models is crucial for subtype identification.
• * Different similarity measures and amalgamation rules impact clustering outcomes.

Purpose of the Study:

• * To compare the performance of 16 hierarchical clustering algorithms.
• * To assess algorithm accuracy in resolving multivariate normal and gamma mixtures.
• * To identify the most effective clustering techniques for subtype discovery.

Main Methods:

• * A 4x4 design comparing four amalgamation rules (single, complete, average, centroid) and four similarity measures (Euclidean, correlation, one-way intraclass, two-way intraclass).
• * Evaluation using Blashfield's (1976) 20 multivariate normal mixtures and Mojena's (1977) 12 multivariate gamma mixtures.
• * Accuracy assessment via kappa as a function of coverage (Edelbrock, 1979) and Rand's statistic, with comparisons to Ward's minimum variance method.

Main Results:

• * Kappa and Rand's statistic yielded identical conclusions regarding algorithm performance.
• * A subset of algorithms showed high accuracy across both mixture types.
• * Specifically, average and centroid linkage using the one-way intraclass correlation, and Ward's technique, were highly accurate.

Conclusions:

• * The one-way intraclass correlation measure is effective in hierarchical clustering for identifying subtypes.
• * These subtypes can differ in profile shape, elevation, and scatter.
• * Average/centroid linkage with one-way intraclass correlation and Ward's method are recommended for mixture model clustering.