Joint Feature Selection and Extraction With Sparse Unsupervised Projection

Area of Science:

• Data Science
• Machine Learning
• Computer Vision

Background:

• Feature selection and feature extraction are key dimensionality reduction strategies with distinct limitations.
• Feature selection offers interpretability but misses implicit sample structures.
• Feature extraction can uncover structures but may sacrifice feature meaning.

Purpose of the Study:

• To propose a novel method combining feature selection and extraction for improved data dimensionality reduction.
• To address the limitations of individual feature selection and extraction techniques.
• To enhance the discovery of implicit structural information within datasets.

Main Methods:

• Introduced joint feature selection and extraction using sparse unsupervised projection (SUP) and graph optimization SUP (GOSUP).
• Incorporated a sparsity constraint on the projection matrix to select relevant features.
• Developed a new algorithm and a 'purification matrix' to eliminate meaningless subspace information.

Main Results:

• The proposed joint method effectively integrates feature selection and extraction.
• Sparsity constraint ensures meaningful feature selection for extraction.
• The 'purification matrix' concept aids in refining sample information in the subspace.
• Empirical results on multiple datasets demonstrate the method's effectiveness.

Conclusions:

• The proposed joint feature selection and extraction method offers a powerful approach to data dimensionality reduction.
• This technique balances feature interpretability with the discovery of underlying data structures.
• The novel algorithm and purification matrix contribute to more effective data subspace representation.