Uncorrelated multilinear principal component analysis for unsupervised multilinear subspace learning.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Classical Principal Component Analysis (PCA) is limited for high-dimensional tensorial data.
• Unsupervised subspace learning requires methods that can handle complex data structures.
• Existing multilinear extensions of PCA have varying effectiveness.

Purpose of the Study:

• To propose an Uncorrelated Multilinear Principal Component Analysis (UMPCA) algorithm.
• To extend the PCA framework for unsupervised subspace learning of tensorial data.
• To develop a method for determining the maximum number of uncorrelated multilinear features.

Main Methods:

• Developed UMPCA as a multilinear extension of PCA.
• Employed successive variance maximization for feature extraction.
• Utilized sequential iterative steps based on the alternating projection method.
• Compared UMPCA against PCA and five state-of-the-art multilinear methods (2DPCA, CSA, TROD, GPCA, MPCA).

Main Results:

• UMPCA successfully captures data variation while producing uncorrelated features.
• The method systematically determines the maximum number of extractable features.
• UMPCA demonstrates effectiveness in unsupervised face and gait recognition tasks.
• Experimental results show UMPCA outperforms baseline PCA and other multilinear extensions.

Conclusions:

• UMPCA is a powerful tool for unsupervised subspace learning of tensorial data.
• The algorithm is particularly effective for dimensionality reduction in recognition tasks.
• UMPCA offers a robust alternative to existing multilinear PCA methods.