Fast Variational Inference for Bayesian Factor Analysis in Single and Multi-Study Settings

Area of Science:

• Statistics
• Computational Biology
• Bioinformatics

Background:

• Factor models are crucial for analyzing high-dimensional data in single and multi-study contexts.
• Bayesian inference for these models typically uses Markov Chain Monte Carlo (MCMC), which struggles with scalability due to increasing data complexity.

Purpose of the Study:

• To develop novel variational inference algorithms for Bayesian latent factor models.
• To address the computational limitations of MCMC methods in high-dimensional settings.

Main Methods:

• Proposed new variational inference algorithms utilizing a multiplicative gamma process shrinkage prior.
• Compared the performance and accuracy of the new algorithms against MCMC implementations.

Main Results:

• The proposed variational inference algorithms provide fast approximate inference.
• These methods require significantly less time and memory compared to MCMC.
• Achieved comparable accuracy in characterizing the data covariance matrix.
• Demonstrated utility in analyzing high-dimensional, multi-study gene expression data from ovarian cancers.

Conclusions:

• The developed variational inference approaches offer efficient and scalable solutions for factor models.
• Facilitates the application of factor models in high-dimensional data analysis.
• An R package, VIMSFA, is available for implementing these methods.