Joint Active Learning with Feature Selection via CUR Matrix Decomposition.

Area of Science:

• Machine Learning
• Data Mining
• Computational Statistics

Background:

• Traditional methods often address sample and feature selection independently.
• Noisy features can negatively impact sample selection, while representative samples aid feature selection.
• Existing approaches struggle with limited or absent supervision.

Purpose of the Study:

• To develop an unsupervised learning framework for simultaneous sample and feature selection.
• To leverage the interleaved nature of these two tasks for improved data representation.
• To address the limitations of existing methods in scenarios with scarce or no labeled data.

Main Methods:

• Proposes a novel framework for joint active learning and feature selection.
• Utilizes CUR matrix decomposition for data reconstruction.
• Employs a one-shot approach, avoiding iterative sample selection for progressive labeling.

Main Results:

• The method effectively approximates the original dataset using selected samples and features.
• Selected samples are highly representative, characterized by the chosen features.
• Experimental results demonstrate superior performance compared to state-of-the-art methods on public datasets.

Conclusions:

• The proposed unsupervised approach offers an effective solution for simultaneous sample and feature selection.
• The one-shot, joint learning framework is particularly advantageous for datasets with limited supervision.
• The method provides a robust and efficient solution to a challenging NP-hard problem.