Bi-Level Spectral Feature Selection.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Traditional unsupervised feature selection (UFS) methods focus solely on the feature level, often neglecting interactions with downstream tasks like classification.
• This limitation can degrade the performance of feature selection, especially for high-dimensional datasets.

Purpose of the Study:

• To propose a novel unsupervised feature selection method that integrates both feature and classification levels.
• To enhance feature selection performance by considering feature relevance for both clustering and classification tasks.

Main Methods:

• Developed a bi-level spectral feature selection (BLSFS) framework.
• Employed spectral clustering to generate pseudo-labels for the classification level.
• Incorporated feature selection at the feature level by preserving data structure using a learned regression matrix.

Main Results:

• The proposed BLSFS method demonstrated superior performance in both clustering and classification tasks across 12 benchmark datasets.
• The integrated bi-level approach effectively captures feature interactions and improves selection accuracy.

Conclusions:

• BLSFS offers a more effective approach to unsupervised feature selection by unifying feature and classification levels.
• The method provides a robust framework for selecting informative features in high-dimensional data for improved downstream task performance.