Consensus Monte Carlo for Random Subsets using Shared Anchors.

Area of Science:

• Computational Statistics
• Machine Learning
• Data Science

Background:

• Existing Bayesian nonparametric models for clustering and feature allocation struggle to scale to big data challenges.
• Efficient algorithms are needed to handle large datasets in statistical modeling and machine learning.

Purpose of the Study:

• To introduce a consensus Monte Carlo algorithm that enables Bayesian nonparametric models to scale for big data applications.
• To validate the algorithm's applicability across diverse priors and sampling models for clustering and feature allocation.

Main Methods:

• Developed a consensus Monte Carlo algorithm designed for scalability with big data.
• The algorithm supports various priors on random subsets (e.g., partitions, latent feature allocation) and sampling models.
• Focused on Dirichlet process mixture models, Indian buffet process priors (binomial sampling), and categorical sampling models.

Main Results:

• The proposed algorithm effectively scales Bayesian nonparametric models for big data.
• Demonstrated successful inference on challenging datasets: MNIST images, pancreatic cancer mutations, and electronic health records (EHR).
• Simulation studies confirmed the algorithm's validity and performance.

Conclusions:

• The consensus Monte Carlo algorithm provides a scalable solution for Bayesian nonparametric big data analysis.
• The method is versatile, applicable to a wide range of clustering and feature allocation problems.
• Successful application to real-world datasets highlights its practical utility in diverse scientific domains.