GPU-powered Shotgun Stochastic Search for Dirichlet process mixtures of Gaussian Graphical Models.

Area of Science:

• Statistics
• Machine Learning
• Computational Biology

Background:

• Gaussian graphical models (GGMs) are widely used for high-dimensional data with sparse dependencies.
• Mixtures of GGMs address population heterogeneity by modeling data from distinct subgroups.
• Dirichlet process mixtures extend these models for flexible clustering and structure discovery.

Purpose of the Study:

• To develop a novel stochastic search algorithm for high-dimensional Dirichlet process mixtures of decomposable GGMs.
• To leverage graphical processing units (GPUs) for accelerating computational efficiency.
• To compare the proposed algorithm against existing Markov chain Monte Carlo (MCMC) methods.

Main Methods:

• A novel stochastic search algorithm is proposed for posterior mode estimation.
• The algorithm is designed to utilize massive thread-parallelization on GPUs.
• Performance is evaluated using simulated datasets and a real gene expression dataset.

Main Results:

• The stochastic search algorithm demonstrates significant improvements in computational speed compared to MCMC.
• The algorithm achieves higher quality posterior mode estimates.
• MCMC methods are found to be impractically slow for certain real-world applications, such as gene expression analysis.

Conclusions:

• The proposed stochastic search algorithm offers a computationally advantageous alternative for analyzing high-dimensional mixture GGMs.
• GPU acceleration significantly enhances the feasibility of applying these models to large datasets.
• This approach provides a practical solution for complex data analysis in fields like bioinformatics.