Generalized Uncorrelated Regression with Adaptive Graph for Unsupervised Feature Selection.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• High-dimensional data often contains redundant features, hindering classification and clustering accuracy.
• Existing unsupervised feature selection methods may fail to address feature redundancy, leading to suboptimal performance.

Purpose of the Study:

• To develop an improved sparse regression model for selecting uncorrelated and discriminative features.
• To enhance unsupervised feature selection by incorporating local data structure and avoiding redundant feature selection.

Main Methods:

• A generalized uncorrelated regression model (GURM) was proposed, utilizing a generalized uncorrelated constraint to seek discriminative features.
• The model leverages the Stiefel manifold to prevent trivial solutions and incorporates a graph regularization term based on maximum entropy (URAFS).
• An efficient algorithm using the generalized powered iteration method was developed for URAFS.

Main Results:

• The proposed URAFS method demonstrated effectiveness and superiority in clustering tasks across eight benchmark datasets.
• Experiments showed that URAFS outperformed seven state-of-the-art methods in feature selection and clustering.

Conclusions:

• The GURM and URAFS models offer a robust approach to unsupervised feature selection by effectively handling feature redundancy and preserving local data geometry.
• The proposed methods provide a significant advancement in the preprocessing of high-dimensional data for clustering and classification tasks.