[Paper Review] GFA: Exploratory Analysis of Multiple Data Sources with Group Factor Analysis

Eemeli Leppäaho, Muhammad Ammad-ud-din+1 more|arXiv (Cornell University)|Nov 3, 2016

Gene expression and cancer classification6 references20 citations

TL;DR

This paper introduces the GFA R package for group factor analysis that enables exploratory analysis of multiple co-occurring data sources through sparse, interpretable factorization. By applying group sparse priors to projection matrices, GFA identifies shared and source-specific latent components, supports missing value prediction, and provides a complete pipeline from preprocessing to robust model interpretation, significantly enhancing multi-source data integration and biclustering.

ABSTRACT

The R package GFA provides a full pipeline for factor analysis of multiple data sources that are represented as matrices with co-occurring samples. It allows learning dependencies between subsets of the data sources, decomposed into latent factors. The package also implements sparse priors for the factorization, providing interpretable biclusters of the multi-source data

Motivation & Objective

• Address the need for a comprehensive, end-to-end pipeline for analyzing multiple co-occurring data sources with complex shared variation and structured noise.
• Overcome limitations of existing tools like CCAGFA and CMF by enabling sparse factorization and providing full analysis workflow support.
• Enable interpretable biclustering of multi-source data through element-wise sparsity priors on both latent factors and loadings.
• Support robust model inference in the presence of multi-modal posteriors common in high-dimensional factor analysis.
• Facilitate model complexity selection via automatic pruning of redundant components while allowing flexible noise modeling.

Proposed method

• Uses Bayesian group factor analysis to model multiple data matrices as low-rank approximations via shared latent factors.
• Applies spike-and-slab priors to induce group sparsity in the projection matrices (W^{(m)}), enabling identification of components shared across subsets of data sources.
• Employs Gibbs sampling for posterior inference, with automatic model complexity selection by initializing with a high number of components (K) and pruning redundant ones.
• Supports missing data imputation by treating NA entries as unobserved and predicting them via posterior predictive distributions.
• Uses a correlation-based robustness analysis (robustComponents()) to identify frequently occurring components across multiple MCMC chains, improving reliability in multi-modal posterior settings.
• Integrates preprocessing (e.g., centering, scaling) and visualization tools (e.g., visualizeComponents()) into a unified analysis pipeline.

Experimental results

Research questions

• RQ1 How can a unified, end-to-end pipeline for multi-source data factorization be designed to support preprocessing, modeling, and interpretation?
• RQ2 To what extent can group sparse priors improve interpretability and discovery of shared or source-specific patterns in multi-source data?
• RQ3 How effective is automatic model complexity selection via component pruning in identifying the true underlying structure without overfitting?
• RQ4 Can robust component identification across multiple MCMC chains improve reliability in high-dimensional, multi-modal posterior settings?
• RQ5 How well does GFA perform in predicting missing values in multi-source data, especially when relationships between sources are complex?

Key findings

• The GFA package successfully implements a complete analysis pipeline for multi-source data, including preprocessing, factorization, missing value prediction, and model interpretation.
• Group sparse priors enable identification of components shared between any subset of data sources, enhancing interpretability and reducing noise in factorization.
• Model complexity selection via automatic pruning works effectively when initial K is sufficiently high, with minimal performance degradation even when K is overestimated.
• On the GDSC cancer cell line dataset, initial K values of 20 or 25 yielded optimal predictive performance (Spearman correlation), with full models remaining competitive.
• The robustComponents() function effectively identifies stable components across multiple MCMC chains, improving reliability in multi-modal posterior scenarios.
• Missing value prediction is accurate and efficient, especially when using fixed projections from a trained model on sequential data batches.