Feature selection for linear SVMs under uncertain data: robust optimization based on difference of convex functions...

Area of Science:

• Machine Learning
• Optimization
• Data Science

Background:

• Real-world datasets often contain inherent uncertainty.
• Feature selection is crucial for building effective linear Support Vector Machines (SVMs).
• Traditional methods struggle with data perturbations.

Purpose of the Study:

• To develop robust feature selection schemes for linear SVMs dealing with uncertain data.
• To address the challenge of the ℓ0-norm in feature selection for uncertain datasets.
• To improve the resilience of SVM models against data perturbations.

Main Methods:

• Utilizing principles of Robust Optimization for handling uncertainty.
• Employing ellipsoidal and box models to represent data uncertainty.
• Applying approximations and Difference of Convex (DC) programming with DC Algorithms (DCA) to manage the ℓ0-norm.

Main Results:

• The proposed robust optimization schemes effectively handle uncertain data.
• The methods overcome the computational difficulties associated with the ℓ0-norm.
• Computational results demonstrate superior performance compared to traditional approaches in immunizing against data perturbations.

Conclusions:

• Robust optimization provides a powerful framework for feature selection on uncertain data.
• The developed methods enhance the stability and reliability of linear SVMs.
• This work offers a significant advancement in handling noisy and uncertain datasets for machine learning models.