Structural learning and integrative decomposition of multi-view data

Area of Science:

• Bioinformatics
• Computational Biology
• Data Science

Background:

• Multi-view data analysis is crucial, with low-rank matrix factorization models widely used.
• Existing models struggle with partially-shared components and determining the number of components.

Purpose of the Study:

• To develop a novel linked component model for multi-view data that incorporates partially-shared structures.
• To enable joint identification of component numbers, overcoming limitations of sequential approaches.

Main Methods:

• Formulation of SLIDE (Structural Learning and Integrative DEcomposition) model.
• Development of model-fitting and selection techniques for joint component number identification.
• Empirical validation on simulated and real-world datasets.

Main Results:

• SLIDE effectively models partially-shared structures in multi-view data.
• The model demonstrates superior performance in signal estimation and component selection.
• Successful application to breast cancer data from The Cancer Genome Atlas.

Conclusions:

• SLIDE offers an advanced approach for multi-view data integration and analysis.
• The methodology provides robust identification of shared, partially-shared, and individual components.
• SLIDE has significant implications for biological data analysis and discovery.