Enriched Pitman-Yor processes

Area of Science:

• Bayesian nonparametrics
• Statistical modeling
• Machine learning

Background:

• Bayesian nonparametrics offers flexible methods for inference and data reduction.
• Discrete prior laws are crucial but existing options like the Dirichlet process have limitations for nested clustering.
• Need for more flexible priors to handle complex data partition structures.

Purpose of the Study:

• Introduce a novel discrete nonparametric prior: the enriched Pitman-Yor process.
• Enhance flexibility in modeling elaborate nested partition structures.
• Provide a unified probabilistic framework for existing Bayesian nonparametric priors.

Main Methods:

• Investigate theoretical properties of the enriched Pitman-Yor process.
• Establish connections with enriched Dirichlet process and normalized random measures.
• Utilize a square-breaking representation and derive closed-form expressions for posterior law and urn schemes.

Main Results:

• The enriched Pitman-Yor process offers higher flexibility for nested clustering.
• Established models like Dirichlet processes and mixture of mixtures are special cases.
• Demonstrated practical utility in a species-sampling ecological problem.

Conclusions:

• The enriched Pitman-Yor process serves as a unified framework for Bayesian nonparametrics.
• Provides a powerful tool for modeling complex hierarchical data structures.
• Applicable to diverse fields requiring flexible clustering and partitioning.