Convex Non-Negative Matrix Factorization With Adaptive Graph for Unsupervised Feature Selection | JoVE Visualize

Area of Science:

Machine Learning
Data Mining
Computer Science

Background:

Unsupervised feature selection (UFS) is crucial for high-dimensional data preprocessing, aiming to reduce redundancy and identify representative features.
Existing UFS methods often employ self-expression or pseudolabel matrix learning independently, with limited success in joint optimization.
Prior literature lacks integrated models that simultaneously leverage self-expression and pseudolabel learning for optimal feature selection.

Purpose of the Study:

To propose a novel unsupervised feature selection method, Convex Non-negative Matrix Factorization with Adaptive Graph Constraint (CNAFS).
To address the limitations of existing methods by integrating pseudolabel matrix learning into the self-expression module within a joint optimization framework.
To enhance feature selection by preserving local manifold structures and reducing data redundancy.

Main Methods:

Developed a novel UFS method, CNAFS, based on convex non-negative matrix factorization with an adaptive graph constraint.
Integrated pseudolabel matrix learning with the self-expression module for simultaneous optimization.
Incorporated two distinct manifold regularizations for the pseudolabel and encoding matrices to maintain local geometrical structure.

Main Results:

The proposed CNAFS method demonstrates effectiveness in selecting the most representative features from high-dimensional data.
Experiments on benchmark datasets validate the superior performance of CNAFS compared to existing UFS techniques.
The joint optimization framework successfully captures data correlations and preserves local manifold structures.

Conclusions:

CNAFS represents a significant advancement in unsupervised feature selection by uniquely integrating pseudolabel matrix learning and self-expression.
The method's ability to preserve local geometric structures and reduce redundancy makes it highly effective for complex datasets.
The study provides a robust and novel approach to UFS, with source code available for reproducibility.