Multilabel Feature Extraction Algorithm via Maximizing Approximated and Symmetrized Normalized Cross-Covariance...

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Multilabel feature extraction (FE) is crucial for managing complex datasets with irrelevant, redundant, or noisy features.
• Traditional methods face challenges in deriving direct eigenvalue problems due to complex formulations of nonlinear dependence measures like the normalized cross-covariance operator.

Purpose of the Study:

• To develop an effective and computationally efficient linear feature extraction method for multilabel classification.
• To address the limitations of existing feature extraction techniques by proposing a novel approximation and symmetrization approach.

Main Methods:

• Approximation of the normalized cross-covariance operator using Moore-Penrose inverse, linear kernel for features, and delta kernel for labels.
• Symmetrization of the approximated operator to create a novel eigenvalue problem for multilabel linear FE.
• Evaluation against seven existing FE techniques using eight performance metrics and three statistical tests across 12 datasets.

Main Results:

• The proposed approximated and symmetrized method yields a novel eigenvalue problem for multilabel linear FE.
• Experimental results demonstrate superior performance of the novel method compared to seven existing FE techniques.
• The method consistently ranked best across eight multilabel classification metrics and three statistical tests.

Conclusions:

• The developed technique offers an effective solution for multilabel feature extraction, enhancing classification performance.
• This novel approach simplifies the process while maintaining or improving accuracy in multilabel learning tasks.
• The findings suggest a significant advancement in the field of feature selection for complex, multilabel datasets.