MULTIVARIATE KERNEL PARTITION PROCESS MIXTURES.

Area of Science:

• Statistics
• Machine Learning
• Biostatistics

Background:

• Mixture models offer flexibility by relaxing parametric assumptions.
• Traditional discrete mixture models assign uniform cluster allocations across parameters in multivariate settings.
• Existing methods can be limited in modeling complex multivariate random effects distributions.

Purpose of the Study:

• To propose a novel kernel partition process (KPP) for flexible, sparse nonparametric modeling of multivariate random effects.
• To develop a KPP that allows varying cluster allocations for different parameters.
• To enhance the modeling of dependence structures in multivariate data.

Main Methods:

• The kernel partition process (KPP) is introduced as a flexible approach for cluster allocation.
• KPP drives a multivariate ordered Chinese restaurant process for dependence modeling.
• An exact block Gibbs sampler is developed for posterior computation without measure truncation.

Main Results:

• The proposed KPP enables spatially-informed clustering of random effects.
• The method demonstrates a highly flexible dependence structure in local clustering.
• The approach was successfully applied to hormone curve data analysis.

Conclusions:

• The kernel partition process (KPP) offers a more flexible alternative to traditional mixture models for multivariate random effects.
• KPP facilitates sparse nonparametric modeling and captures complex dependence structures.
• The developed methods are applicable to real-world data, such as hormone curve analysis, and can be extended for classification tasks.