Semisupervised Feature Selection With Sparse Discriminative Least Squares Regression.

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Selecting informative features is critical in the era of big data.
• Supervised learning methods require substantial labeled data, which is often costly to obtain.
• Semisupervised learning offers a promising alternative by leveraging both labeled and unlabeled data.

Purpose of the Study:

• To propose a novel sparse discriminative semisupervised feature selection (SDSSFS) method.
• To enhance feature selection by exploiting both labeled and unlabeled data effectively.
• To achieve a more discriminative and sparse feature subset.

Main Methods:

• Extending the ϵ-dragging technique from supervised to semisupervised learning to increase inter-class distances.
• Employing the flexible l_2,p norm for implicit regularization to promote sparsity.
• Developing an iterative algorithm for simultaneous learning of regression coefficients, ϵ-dragging matrix, and predicting unknown labels.

Main Results:

• The proposed SDSSFS method demonstrated superior performance across ten real-world datasets.
• The method successfully identified informative features while handling limited labeled data.
• Achieved enhanced discriminative power and sparsity in the selected feature subsets.

Conclusions:

• The SDSSFS method provides an effective approach for feature selection in semisupervised learning scenarios.
• The integration of ϵ-dragging and l_2,p norm regularization offers significant advantages.
• The proposed iterative algorithm efficiently solves the complex optimization problem.