A Dirichlet process mixture of generalized Dirichlet distributions for proportional data modeling

Area of Science:

• Machine Learning
• Statistical Modeling
• Nonparametric Bayesian Analysis

Background:

• Clustering algorithms are essential for data analysis and pattern recognition.
• Modeling proportional data often requires specialized distributions.
• Finite mixture models necessitate pre-specifying the number of components, limiting flexibility.

Purpose of the Study:

• To propose an infinite clustering algorithm for proportional data.
• To extend finite generalized Dirichlet mixture models to the infinite case using nonparametric Bayesian methods.
• To develop a clustering approach that automatically estimates the number of mixture components.

Main Methods:

• Utilizing Dirichlet processes and generalized Dirichlet distributions for flexible proportional data modeling.
• Employing nonparametric Bayesian analysis for an infinite mixture model extension.
• Implementing a Gibbs sampler for estimating the posterior distribution of clusterings.

Main Results:

• The proposed algorithm successfully clusters proportional data without prior component specification.
• Demonstrated superior and more robust performance compared to classic finite mixture models.
• Validated through applications in real-data classification and image database categorization.

Conclusions:

• Infinite mixture models offer a powerful and robust alternative for clustering proportional data.
• The nonparametric Bayesian approach provides a principled way to estimate the number of clusters.
• This algorithm enhances data analysis capabilities in classification and image categorization tasks.