Principal component analysis based on l1-norm maximization.

Area of Science:

• Data Science
• Machine Learning
• Statistics

Background:

• Principal Component Analysis (PCA) is a widely used dimensionality reduction technique.
• Conventional PCA relies on L2-norm, making it sensitive to outliers in datasets.
• Robustness to outliers is crucial for reliable data analysis in various fields.

Purpose of the Study:

• To propose a novel L1-norm optimization technique for Principal Component Analysis (PCA).
• To enhance the robustness of PCA against outliers.
• To develop a rotation-invariant PCA method that is simple and easy to implement.

Main Methods:

• Developed a new L1-norm optimization technique for PCA.
• Demonstrated the outlier robustness of the L1-norm approach compared to L2-norm.
• Proved that the proposed method finds a locally maximal solution.
• Applied the method to multiple datasets for performance evaluation.

Main Results:

• The proposed L1-norm PCA method exhibits superior robustness to outliers.
• The technique is invariant to data rotations.
• The L1-norm optimization is intuitive, simple, and easy to implement.
• Empirical results show competitive or improved performance compared to conventional methods.

Conclusions:

• The L1-norm optimization technique offers a robust and effective alternative to conventional PCA.
• This method provides a valuable tool for data analysis where outliers are a concern.
• The simplicity and rotation invariance make it broadly applicable in machine learning and statistics.