Robust extraction of local structures by the minimum beta-divergence method.

Area of Science:

• Machine Learning
• Statistical Modeling
• Data Analysis

Background:

• Principal Component Analysis (PCA) is crucial for dimensionality reduction.
• Heterogeneous data often requires specialized methods for accurate analysis.
• Existing methods may struggle with identifying local structures in mixed data models.

Purpose of the Study:

• To develop a highly robust learning algorithm for exploring local PCA structures.
• To address the challenge of analyzing data that follows multiple heterogeneous PCA models.
• To provide a method that can effectively identify PCA patterns within specific data clusters.

Main Methods:

• The proposed method minimizes beta-divergence for robustness.
• It utilizes an initial shifting parameter and a tuning parameter beta to search for local PCA structures.
• The algorithm identifies a data cluster's PCA structure by treating data from other clusters as outliers.

Main Results:

• The algorithm successfully detects local PCA structures in simulations.
• It demonstrates robustness in identifying patterns within specific data clusters.
• The method's performance is validated through simulation studies.

Conclusions:

• The developed algorithm offers a robust approach to analyzing heterogeneous data with local PCA structures.
• It provides a viable alternative to methods like finite mixture models for such data.
• The study highlights the importance of parameter selection for optimal performance.