Robust Principal Component Analysis via Joint Reconstruction and Projection.

Area of Science:

• Data Science
• Machine Learning
• Statistics

Background:

• Principal Component Analysis (PCA) is a standard unsupervised dimensionality reduction technique.
• Standard PCA is sensitive to outliers due to its reliance on the squared l2-norm distance metric.
• Existing robust PCA methods often focus on either reconstruction error or projection variance, neglecting their simultaneous optimization.

Purpose of the Study:

• To develop a novel robust principal component analysis (RPCA) method that addresses the limitations of standard PCA.
• To improve the robustness of PCA against outliers by considering both reconstruction error and projection variance.
• To introduce anomaly detection capabilities into the PCA framework.

Main Methods:

• Propose Robust Principal Component Analysis via Joint Reconstruction and Projection (RPCA-RP).
• Incorporate a discrete weighting mechanism to differentiate between normal data and outliers.
• Develop an iterative algorithm for solving the RPCA-RP optimization problem.
• Conduct theoretical analysis of the proposed method.

Main Results:

• RPCA-RP effectively combines reconstruction error and projection variance for enhanced data analysis.
• The discrete weighting scheme successfully identifies and mitigates the impact of outliers.
• The method demonstrates significant robustness and improved performance on various real-world and large-scale datasets.
• Unexpectedly, RPCA-RP exhibits strong anomaly detection capabilities.

Conclusions:

• RPCA-RP offers a superior and more robust alternative to traditional PCA, particularly in the presence of outliers.
• The joint optimization of reconstruction and projection, along with discrete weighting, is key to the method's effectiveness.
• The discovered anomaly detection capability adds significant value to the proposed technique, broadening its applicability.