Correlation metric for generalized feature extraction.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Conventional feature extraction methods like linear and kernel-based approaches have limitations.
• Normalized data, common in preprocessing, requires specialized feature extraction techniques.
• Euclidean distance may not be optimal for classification compared to correlation metrics.

Purpose of the Study:

• To propose a generalized feature extraction framework using Graph Embedding.
• To introduce two novel algorithms, Correlation Embedding Analysis (CEA) and Correlational Principal Component Analysis (CPCA).
• To demonstrate the effectiveness of these new algorithms for visual classification tasks.

Main Methods:

• Developed a generalized feature extraction formulation based on the Graph Embedding framework.
• Introduced Correlation Embedding Analysis (CEA) for correlational mapping and discriminating analysis.
• Introduced Correlational Principal Component Analysis (CPCA) generalizing Principal Component Analysis (PCA) for hyperspherical data.

Main Results:

• CEA maps high-dimensional hyperspherical data to low-dimensional hyperspheres, preserving neighbor relations.
• CPCA extends PCA for data distributed on high-dimensional hyperspheres.
• Both algorithms are tailored for normalized data and leverage correlation metrics for improved classification.

Conclusions:

• The proposed Graph Embedding framework and its algorithms (CEA, CPCA) offer effective feature extraction.
• Correlation metric-based methods show advantages over Euclidean distance for classification.
• Experimental results validate the superior performance of the proposed algorithms in visual classification.