Joint Anchor Graph Embedding and Discrete Feature Scoring for Unsupervised Feature Selection.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Unsupervised feature selection (UFS) methods often assume intrinsic data relationships exist in a low-dimensional subspace.
• Existing UFS methods struggle with computational complexity and local optima, limiting their use on large datasets.

Purpose of the Study:

• To develop a novel UFS method that overcomes the limitations of existing approaches for large-scale datasets.
• To improve the optimality of selected features and enhance the performance of downstream tasks like clustering and image segmentation.

Main Methods:

• Proposes a novel anchor graph embedding paradigm for efficient extraction of local neighborhood relationships.
• Introduces a discrete feature scoring mechanism with orthogonal l2,0-norm constraints to enhance feature distinction and avoid local optima.
• Develops an efficient optimization algorithm to solve the NP-hard problem and obtain a closed-form solution for the transformation matrix.

Main Results:

• The proposed UFS method demonstrates effectiveness and efficiency in extensive experiments.
• Achieves superior performance in clustering and image segmentation tasks compared to state-of-the-art approaches.
• Reduces computational complexity of graph construction to be linear in the number of data points.

Conclusions:

• The novel UFS method offers a computationally efficient and effective solution for feature selection on large-scale datasets.
• The anchor graph embedding and discrete feature scoring mechanism significantly improve feature selection optimality and downstream task performance.
• This approach provides a promising alternative for UFS in machine learning and computer vision applications.