Uncertainty-Aware PCA Revisited.

Area of Science:

• Statistics
• Data Visualization
• Machine Learning

Background:

• Principal Component Analysis (PCA) is a key dimensionality reduction technique.
• Existing PCA methods do not account for uncertainty in high-dimensional data points.
• Small data uncertainties can lead to significant projection uncertainties in standard PCA.

Purpose of the Study:

• To develop a method for quantifying uncertainty in PCA when data points have Gaussian uncertainty.
• To propose a visualization tool to assess the suitability of uncertainty-aware PCA techniques.

Main Methods:

• Derivation of a closed-form expression to quantify eigenvector uncertainty.
• Development of a 3D glyph for visualizing eigenvector uncertainty.
• Application and testing on various datasets.

Main Results:

• Demonstrated that data uncertainty propagates to eigenvector uncertainty in PCA.
• Provided a closed-form solution for eigenvector uncertainty quantification.
• Introduced a 3D glyph to aid in selecting appropriate uncertainty-aware PCA methods.

Conclusions:

• The proposed method effectively quantifies eigenvector uncertainty in PCA under Gaussian data uncertainty.
• The 3D glyph assists in choosing between standard and sampling-based uncertainty-aware PCA approaches.
• This work enhances the reliability of PCA for uncertain high-dimensional data.