Two-Dimensional Unsupervised Feature Selection via Sparse Feature Filter.

Area of Science:

• Computer Science
• Machine Learning
• Image Analysis

Background:

• Unsupervised feature selection is crucial for data learning but remains challenging.
• Current 2-D methods utilize image structure but require hyperparameter tuning.
• This limits the applicability of unsupervised algorithms in image analysis.

Purpose of the Study:

• To design a novel feature filter for unsupervised 2-D feature selection.
• To overcome limitations of existing methods, such as hyperparameter dependency.
• To effectively estimate image feature weights for improved data learning.

Main Methods:

• Developed a feature filter to estimate image feature weights.
• Theoretically demonstrated the filter's equivalence to sparse regularization.
• Implemented two strategies: multiple feature filters and single common feature filter.

Main Results:

• The proposed feature filter effectively estimates feature weights.
• Theoretical analysis confirms its role similar to sparse regularization.
• Achieved superior performance on seven benchmark datasets compared to state-of-the-art methods.

Conclusions:

• The novel unsupervised 2-D weight-based feature selection methods offer superior performance.
• The feature filter approach eliminates the need for additional hyperparameters.
• This work advances unsupervised learning for image analysis applications.