D-CCA: A Decomposition-based Canonical Correlation Analysis for High-Dimensional Datasets.

Area of Science:

• Multivariate Statistics
• Bioinformatics
• Machine Learning

Background:

• Joint analysis of high-dimensional datasets often involves matrix decomposition into common, distinctive, and noise components.
• Existing methods may not adequately enforce orthogonality between distinctive matrices, potentially leaving shared information.
• Canonical Correlation Analysis (CCA) is a standard technique for exploring relationships between two datasets.

Purpose of the Study:

• To propose a novel decomposition-based canonical correlation analysis (D-CCA) method for joint analysis of high-dimensional data.
• To define common and distinctive matrices in the space of random variables, ensuring orthogonality between distinctive matrices.
• To improve upon existing decomposition methods by better separating shared and individual information.

Main Methods:

• Proposed decomposition-based canonical correlation analysis (D-CCA).
• Defined common and distinctive matrices in the space of random variables.
• Constructed orthogonal relationships between distinctive matrices to prevent information leakage.

Main Results:

• The proposed D-CCA method provides consistent estimators for common and distinctive matrices.
• D-CCA demonstrated superior performance compared to state-of-the-art methods on simulated data.
• Applied D-CCA to The Cancer Genome Atlas breast cancer data, showing effective real-world applicability.

Conclusions:

• D-CCA offers a robust framework for joint analysis of high-dimensional datasets by enhancing information separation.
• The method's focus on orthogonality between distinctive matrices is crucial for accurate joint analysis.
• D-CCA represents a significant advancement over traditional CCA and existing decomposition techniques.