Generalized Matrix Factorization: efficient algorithms for fitting generalized linear latent variable models to large...

Area of Science:

• Statistical modeling
• Machine learning
• Data analysis

Background:

• Latent variables explain correlations in multivariate data across psychology, ecology, and medicine.
• Classical methods like factor analysis work for Gaussian data but not non-Gaussian responses.
• Current Generalized Linear Latent Variable models (GLLVMs) are computationally intensive and do not scale to large datasets.

Purpose of the Study:

• To develop a computationally efficient and stable algorithm for fitting GLLVMs to high-dimensional datasets.
• To enable the analysis of large-scale ecological, medical, and psychological data previously intractable with existing methods.

Main Methods:

• Approximation of GLLVMs using penalized quasi-likelihood.
• Application of Newton method and Fisher scoring for parameter estimation.
• Development of an easy-to-use implementation of the fitting algorithm.

Main Results:

• The proposed method is significantly faster and more stable than existing algorithms for GLLVMs.
• The approach successfully fits GLLVMs to datasets with thousands of observational units and responses.
• Analysis of a large ecological dataset revealed that a few latent factors explain most of the species variability.

Conclusions:

• The new penalized quasi-likelihood approach provides a scalable solution for fitting GLLVMs.
• This advancement opens new possibilities for analyzing complex, high-dimensional data in various scientific domains.
• The readily available implementation facilitates broader adoption and application of advanced latent variable modeling.