Disentangling Interpretable Factors with Supervised Independent Subspace Principal Component Analysis.

Area of Science:

• Machine Learning
• Data Science
• Bioinformatics

Background:

• High-dimensional data representation is crucial for machine learning success.
• Current methods struggle to balance interpretability with complex data modeling.
• Linear methods are limited in multi-subspace learning, while deep learning lacks transparency.

Purpose of the Study:

• Introduce Supervised Independent Subspace Principal Component Analysis (sisPCA) for multi-subspace learning.
• Develop a method that incorporates supervision and ensures subspace disentanglement.
• Enhance the interpretability of machine learning models for high-dimensional data.

Main Methods:

• Extension of Principal Component Analysis (PCA) for multi-subspace learning.
• Utilizes the Hilbert-Schmidt Independence Criterion (HSIC) for supervision and disentanglement.
• Demonstrates connections with autoencoders and regularized linear regression.

Main Results:

• Successfully identifies and separates hidden data structures in complex datasets.
• Applied to breast cancer diagnosis, DNA methylation analysis, and single-cell malaria infection studies.
• Revealed distinct functional pathways associated with malaria colonization.

Conclusions:

• sisPCA offers an explainable approach to representation learning in high-dimensional data.
• The method facilitates the discovery of biologically relevant patterns.
• Highlights the importance of interpretable models in scientific discovery.