Scalable Random Feature Latent Variable Models.

Area of Science:

• Machine Learning
• Statistical Modeling
• Data Science

Background:

• Random Feature Latent Variable Models (RFLVMs) excel at uncovering structure in complex, high-dimensional data.
• However, traditional RFLVMs face scalability limitations due to Monte Carlo sampling methods, hindering large-scale applications.

Purpose of the Study:

• To develop a scalable RFLVM framework overcoming Monte Carlo sampling limitations.
• To enable efficient analysis of high-dimensional, non-Gaussian datasets.

Main Methods:

• Developed a scalable RFLVM (SRFLVM) using variational Bayesian inference (VBI).
• Addressed VBI challenges with Dirichlet process (DP) mixing weights using stick-breaking construction.
• Introduced block coordinate descent variational inference (BCD-VI) for efficient optimization of high-dimensional variational parameters.

Main Results:

• SRFLVM demonstrates superior scalability and computational efficiency compared to traditional RFLVMs.
• Achieved state-of-the-art performance in latent representation learning and missing data imputation.
• Outperformed deep generative models and other latent variable models on benchmark datasets.

Conclusions:

• The proposed SRFLVM framework effectively scales RFLVMs for large-scale applications.
• SRFLVM offers a powerful and efficient alternative for analyzing complex, high-dimensional data.
• This work advances the field of latent variable modeling with a focus on practical, scalable solutions.