Avoiding Optimal Mean ℓ_2,1-Norm Maximization-Based Robust PCA for Reconstruction.

Area of Science:

• Machine Learning
• Data Science
• Dimensionality Reduction

Background:

• Robust Principal Component Analysis (PCA) is crucial for high-dimensional data with outliers.
• Existing robust PCA methods often incorrectly assume a zero mean or use suboptimal means.
• Traditional PCA's mean assumption is limited to squared L2-norm, not general robust PCA.

Purpose of the Study:

• To reformulate the objective of conventional PCA for improved robustness.
• To develop a method that is invariant to rotation and avoids optimal mean calculation.
• To connect the reformulated objective to reconstruction error minimization.

Main Methods:

• Reformulated PCA objective by maximizing the sum of projected differences between instances using the L1-norm.
• Developed an efficient iterative reweighting optimization algorithm to handle nonsmooth objectives and outliers.
• Theoretically analyzed the convergence and computational complexity of the proposed algorithm.

Main Results:

• The proposed method demonstrates robustness to outliers and rotational invariance.
• The reformulated objective naturally avoids the need for optimal mean calculation and centered data assumptions.
• Experimental results on benchmark datasets show superior performance compared to existing methods.

Conclusions:

• The novel robust PCA method effectively addresses limitations of existing techniques.
• The approach offers improved accuracy and applicability for high-dimensional data with outliers.
• The method provides a theoretically sound and computationally efficient solution.