Unified Simultaneous Clustering and Feature Selection for Unlabeled and Labeled Data

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Feature selection is crucial for dimensionality reduction and improving model performance.
• Existing methods often struggle with both supervised and unsupervised learning scenarios simultaneously.
• The need for robust feature selection that can leverage latent data structures is evident.

Purpose of the Study:

• To propose a novel unified simultaneous clustering feature selection (USCFS) method.
• To develop a regularized regression framework capable of selecting discriminative features from labeled or unlabeled data.
• To enable effective feature selection without relying on affinity graph-based methods.

Main Methods:

• Formulation of a regularized regression model with a novel target matrix.
• Utilizing L-norm regularization for discriminative feature projection.
• Employing orthogonal basis clustering to capture latent cluster centers for unsupervised selection.
• Leveraging ground-truth labels to guide latent class label discovery for supervised selection.

Main Results:

• The proposed USCFS method effectively selects discriminative features in both supervised and unsupervised scenarios.
• Experimental results show superior performance compared to state-of-the-art methods on diverse real-world datasets.
• The method successfully captures label-like information from projected data for unsupervised tasks.

Conclusions:

• USCFS provides an effective and unified approach to feature selection.
• The method demonstrates robustness and high performance across various datasets and learning paradigms.
• USCFS offers a significant advancement in simultaneous clustering and feature selection techniques.