Extended Stochastic Gradient MCMC for Large-Scale Bayesian Variable Selection.

Area of Science:

• Computational Statistics
• Bayesian Inference
• Machine Learning

Background:

• Stochastic gradient Markov chain Monte Carlo (MCMC) algorithms are vital for Bayesian computing with big data.
• Existing MCMC methods face limitations, requiring fixed parameter dimensions and differentiable log-posterior densities.
• These constraints restrict their application to a narrow range of complex problems.

Purpose of the Study:

• To develop an extended stochastic gradient MCMC algorithm.
• To broaden the applicability of MCMC methods to more general large-scale Bayesian computing.
• To address challenges like dimension jumping and missing data in Bayesian analysis.

Main Methods:

• Introduction of latent variables to extend the applicability of stochastic gradient MCMC.
• Development of a novel algorithm capable of handling variable parameter spaces.
• Implementation and testing on large-scale Bayesian computing problems.

Main Results:

• The proposed extended MCMC algorithm demonstrates high scalability for big data.
• Numerical studies confirm significantly improved efficiency compared to traditional MCMC algorithms.
• The algorithm effectively handles complex scenarios including dimension jumping and missing data.

Conclusions:

• The extended stochastic gradient MCMC algorithm significantly advances Bayesian computing for big data.
• This approach alleviates previous computational burdens associated with Bayesian methods.
• The method offers a more versatile and efficient solution for modern statistical challenges.