Unsupervised feature selection via adaptive autoencoder with redundancy control.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• High-dimensional unlabeled data presents challenges in dimensionality reduction.
• Unsupervised feature selection is crucial for efficient data analysis without prior labels.
• Existing methods may not effectively control redundancy or optimize network structure simultaneously.

Purpose of the Study:

• To introduce a novel unsupervised feature selection method called Adaptive Autoencoder with Redundancy Control (AARC).
• To integrate feature selection, network structure optimization, and redundancy control into a unified framework.
• To enhance the efficiency and effectiveness of dimensionality reduction for unlabeled high-dimensional datasets.

Main Methods:

• Developed a novel Adaptive Autoencoder with Redundancy Control (AARC).
• Incorporated Group Lasso penalties for feature selection and compact network structure determination.
• Added a feature dependency penalty (e.g., Pearson correlation, mutual information) to control redundancy.
• Utilized adaptive parameters and a smoothing function for differentiable optimization.

Main Results:

• Ablation studies validated AARC's redundancy control and structure optimization capabilities.
• AARC demonstrated superior efficiency compared to nine state-of-the-art unsupervised feature selection methods.
• The proposed method effectively reduces data dimensionality while preserving important features.

Conclusions:

• AARC offers an effective and integrated approach to unsupervised feature selection.
• The method successfully balances feature selection, network simplification, and redundancy minimization.
• AARC shows significant promise for applications involving high-dimensional, unlabeled data analysis.