Robust principal component analysis based on maximum correntropy criterion.

Area of Science:

• Machine Learning
• Data Analysis
• Statistical Modeling

Background:

• Principal Component Analysis (PCA) is a widely used dimensionality reduction technique.
• Standard PCA is sensitive to outliers and assumes zero-mean data.
• Existing robust PCA methods may lack theoretical grounding or struggle with nonlinear data.

Purpose of the Study:

• To develop a novel rotational-invariant PCA method robust to outliers.
• To enhance PCA's applicability to datasets with non-zero means and nonlinear structures.
• To provide a theoretically sound alternative to MSE-based PCA.

Main Methods:

• Utilizing the Maximum Correntropy Criterion (MCC) for robust estimation.
• Employing a half-quadratic optimization algorithm to solve the correntropy objective.
• Integrating kernel techniques to address nonlinear data distributions.

Main Results:

• The proposed MCC-based PCA demonstrates robustness against outliers.
• The method effectively estimates data mean without prior assumptions.
• It yields principal eigenvectors of a robust covariance matrix.
• Performance surpasses L(1) norm-based robust PCA in the presence of outliers.

Conclusions:

• The MCC-based rotational-invariant PCA offers superior robustness and flexibility.
• This approach provides a theoretically solid and efficient alternative for data analysis.
• The integration of kernel methods expands its utility for complex datasets.