Spherical Rotation Dimension Reduction with Geometric Loss Functions.

Area of Science:

• Data Science
• Computational Biology
• Geometric Data Analysis

Background:

• Modern datasets frequently display high dimensionality.
• Underlying low-dimensional manifolds contain critical geometric structures for analysis.
• Cyclical biological processes, such as the cell cycle, can be modeled as spherical manifolds.

Purpose of the Study:

• To introduce Spherical Rotation Component Analysis (SRCA), a novel nonlinear dimension reduction method.
• To incorporate geometric information for improved approximation of low-dimensional manifolds.
• To provide a versatile method applicable to high-dimensional data and small sample sizes.

Main Methods:

• SRCA employs spheres or ellipsoids to represent data.
• It achieves a low-rank spherical data representation with theoretical guarantees.
• The method effectively retains the intrinsic geometric structure during dimensionality reduction.

Main Results:

• SRCA demonstrates superior performance in approximating manifolds compared to existing methods.
• The technique successfully preserves inherent geometric structures in datasets.
• Simulations and application to human cell cycle data validate SRCA's effectiveness.

Conclusions:

• SRCA offers a robust approach for analyzing high-dimensional datasets with underlying geometric structures.
• The method is particularly well-suited for cyclical data, such as cell cycle measurements.
• SRCA advances the field of nonlinear dimension reduction by integrating geometric principles.