Unsupervised Adaptive Feature Selection With Binary Hashing.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Unsupervised feature selection methods often lack label guidance, leading to information loss and semantic limitations, especially with multi-label data.
• Existing approaches struggle with datasets annotated with multiple labels, hindering effective feature selection.
• Real-world data like images and videos frequently possess multiple annotations, necessitating advanced feature selection techniques.

Purpose of the Study:

• To propose a novel Unsupervised Adaptive Feature Selection with Binary Hashing (UAFS-BH) model for effective feature selection in unsupervised learning.
• To address the limitations of existing methods by incorporating weakly-supervised multi-labels derived from binary hash codes.
• To enhance feature selection by adaptively learning multi-labels and modeling intrinsic data structure.

Main Methods:

• The UAFS-BH model learns binary hash codes as weakly-supervised multi-labels to guide feature selection.
• Weakly-supervised multi-labels are automatically generated by imposing binary hash constraints on spectral embedding.
• The number of multi-labels is adaptively determined, and a dynamic similarity graph is constructed to capture data structure.

Main Results:

• The proposed UAFS-BH model demonstrates state-of-the-art performance on single-view feature selection tasks.
• The Multi-view Feature Selection with Binary Hashing (MVFS-BH) extension achieves superior results in multi-view feature selection.
• Extensive experiments on benchmark datasets validate the effectiveness of the proposed approach.

Conclusions:

• The UAFS-BH model offers a significant advancement in unsupervised feature selection for multi-label data.
• The adaptive learning of weak multi-labels and dynamic similarity graph construction are key to the method's success.
• The developed approach provides a robust solution for both single-view and multi-view feature selection problems.