SPARSE ENCODING FOR MORE-INTERPRETABLE FEATURE-SELECTING REPRESENTATIONS IN PROBABILISTIC MATRIX FACTORIZATION.

Area of Science:

• Medical Informatics
• Computational Biology
• Statistical Modeling

Background:

• Dimensionality reduction is crucial for interpretable count data analysis.
• Sparse probabilistic non-negative matrix factorization (NMF) methods like hierarchical Poisson matrix factorization (HPF) offer interpretability through sparse decoding.
• However, HPF lacks encoder sparsity, limiting its ability to define factor-feature relationships.

Purpose of the Study:

• To address the deficiency of encoder sparsity in HPF.
• To develop a method that enforces encoder sparsity for improved interpretability and feature selection.
• To demonstrate the practical utility of encoder sparsity in medical informatics applications.

Main Methods:

• Self-consistently enforcing encoder sparsity within the HPF framework.
• Utilizing a generalized additive model (GAM) to relate representation coordinates to original data features.
• Applying the enhanced method to simulated data and a real-world dataset of Medicare patient comorbidities.

Main Results:

• The proposed method successfully enforces encoder sparsity, unlike standard HPF.
• The method enables the identification of relevant features for each representation coordinate, facilitating feature selection.
• Demonstrated practical application in representing inpatient comorbidities for Medicare patients.

Conclusions:

• Enforcing encoder sparsity in HPF significantly enhances model interpretability and feature selection capabilities.
• The integration of GAMs provides a robust framework for achieving encoder sparsity.
• This approach offers a valuable tool for analyzing high-dimensional count data in medical informatics and beyond.