$\ell _{2,p}$ -Norm Based PCA for Image Recognition.

Area of Science:

• Machine Learning
• Data Science
• Robust Statistics

Background:

• Principal Component Analysis (PCA) is widely used for dimensionality reduction.
• Existing robust PCA methods often focus on maximizing variance, not minimizing reconstruction error.
• Many robust PCA techniques lack rotational invariance, a key property of standard PCA.

Purpose of the Study:

• To propose a generalized robust metric learning approach for PCA, termed ℓ₂‚ₚ-PCA.
• To address limitations of existing ℓ₁-norm-based PCA methods, specifically regarding reconstruction error minimization and rotational invariance.
• To develop a robust PCA method that retains desirable properties of standard PCA.

Main Methods:

• Introduced ℓ₂‚ₚ-PCA, utilizing the ℓ₂‚ₚ-norm as the distance metric for reconstruction error.
• Developed a novel iterative algorithm for efficient computation of the ℓ₂‚ₚ-PCA solution.
• Evaluated the method against standard PCA and several ℓ₁-norm-based robust PCA algorithms.

Main Results:

• The proposed ℓ₂‚ₚ-PCA method demonstrates robustness to outliers.
• ℓ₂‚ₚ-PCA solutions yield principal eigenvectors of a robust covariance matrix.
• The low-dimensional representations obtained by ℓ₂‚ₚ-PCA exhibit rotational invariance.
• Experimental results show superior effectiveness and robustness compared to PCA, PCA-L1 greedy, PCA-L1 nongreedy, and HQ-PCA.

Conclusions:

• ℓ₂‚ₚ-PCA offers a robust and effective alternative for dimensionality reduction.
• The method successfully minimizes reconstruction error while preserving PCA's beneficial properties.
• ℓ₂‚ₚ-PCA provides a significant advancement over existing robust PCA techniques, particularly in handling outliers and maintaining rotational invariance.