The Influence Function of Principal Component Analysis by Self-Organizing Rule

Area of Science:

• Machine Learning
• Statistics
• Data Analysis

Background:

• Principal Component Analysis (PCA) is a fundamental dimensionality reduction technique.
• Neural network approaches offer alternative methods for PCA.
• Robustness against outliers is crucial for reliable PCA results.

Purpose of the Study:

• To investigate the robustness of a self-organizing rule-based neural network algorithm for PCA.
• To analyze the algorithm's performance specifically in the presence of outliers.
• To develop a statistic for assessing data influence in PCA.

Main Methods:

• Utilized the theory of influence functions to analyze outlier sensitivity.
• Derived an explicit form for the influence function of the principal component vector.
• Proposed a novel statistic derived from the self-organizing rule.

Main Results:

• The neural network PCA algorithm demonstrates robustness against outliers in directions orthogonal to the principal component vector.
• The derived influence function provides a theoretical basis for this robustness.
• A new statistic effectively assesses the influence of individual data points.

Conclusions:

• The self-organizing rule-based neural network approach for PCA exhibits significant robustness to outliers.
• The developed influence function and statistic offer valuable tools for robust PCA.
• This method enhances the reliability of PCA in datasets with potential outliers.