Optimal Bayesian estimators for latent variable cluster models.

Area of Science:

• Statistics
• Machine Learning
• Data Mining

Background:

• Cluster analysis aims to group similar individuals or items.
• Bayesian methods offer probabilistic clustering but lack scalable interpretation tools for latent allocation variables.
• Existing methods struggle with categorical clustering variables and determining the optimal number of groups.

Purpose of the Study:

• To develop a scalable Bayesian decision-theoretic framework for cluster analysis.
• To propose a fast, context-independent greedy algorithm for optimal cluster allocation.
• To simultaneously solve clustering and model-choice problems by automatically selecting the optimal number of groups.

Main Methods:

• Utilized a Bayesian decision-theoretic approach to define an optimality criterion for clusterings.
• Developed a fast and context-independent greedy algorithm for finding optimal allocations.
• Incorporated various loss functions to compare and evaluate different partitions.
• Applied the approach to Gaussian mixtures, stochastic block models, and latent block models.

Main Results:

• The proposed greedy algorithm efficiently finds optimal cluster allocations.
• The method automatically selects the optimal number of groups, addressing model-choice uncertainty.
• Demonstrated effectiveness across diverse clustering models and datasets (artificial and real).

Conclusions:

• The developed Bayesian framework provides a robust and scalable solution for interpreting clustering results.
• The automatic selection of the number of clusters simplifies model selection in cluster analysis.
• This approach offers a versatile tool for various applications requiring probabilistic partitioning.