Fast Sparse Discriminative K-Means for Unsupervised Feature Selection.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Embedded feature selection methods leverage pseudolabel matrices for projection matrix learning.
• Continuous pseudolabel matrices derived from spectral analysis can deviate from actual data distributions.
• Existing methods may require complex constraints, complicating optimization.

Purpose of the Study:

• To develop an efficient feature selection framework addressing limitations of continuous pseudolabel matrices.
• To introduce a novel method, Fast Sparse Discriminative K-means (FSDK), for robust feature selection.
• To enhance the optimization process by simplifying constraints.

Main Methods:

• The proposed FSDK framework integrates principles from Least-Squares Regression (LSR) and Discriminative K-means (DisK-means).
• A weighted pseudolabel matrix with discrete properties is introduced to prevent trivial solutions in unsupervised LSR.
• An l2,p-norm regularizer is employed to enforce row sparsity in the selection matrix, offering flexibility with parameter p.

Main Results:

• The FSDK model effectively optimizes a sparse regression problem by combining DisK-means and the l2,p-norm regularizer.
• The computational complexity of FSDK is linearly correlated with the number of samples, enabling rapid processing of large-scale datasets.
• Extensive experiments across diverse datasets demonstrate the superior effectiveness and efficiency of the FSDK approach.

Conclusions:

• FSDK offers a significant advancement in embedded feature selection, providing a more accurate and efficient alternative.
• The discrete pseudolabel matrix and flexible sparsity regularization contribute to the model's robustness and performance.
• The scalability of FSDK makes it well-suited for analyzing large and complex datasets in various domains.