Robust capped L1-norm twin support vector machine.

Area of Science:

• Machine Learning
• Computer Science
• Data Mining

Background:

• Traditional Twin Support Vector Machines (TWSVM) are effective for binary classification.
• TWSVM's reliance on squared L2-norm distance makes it susceptible to outliers, compromising robustness.
• Outliers can disproportionately influence classification models, leading to reduced accuracy.

Purpose of the Study:

• To develop a robust classification method that addresses the outlier sensitivity of TWSVM.
• To introduce the capped L1-norm twin support vector machine (CTWSVM) as a more resilient alternative.
• To enhance the reliability of binary classification in the presence of noisy data.

Main Methods:

• Proposed a novel robust capped L1-norm twin support vector machine (CTWSVM).
• Developed an effective iterative algorithm to solve the CTWSVM optimization problems.
• Conducted theoretical analysis on local optimum existence and algorithm convergence.

Main Results:

• CTWSVM demonstrates superior robustness compared to classical TWSVM when dealing with outliers.
• Extensive experiments on artificial and UCI datasets validate the feasibility and effectiveness of CTWSVM.
• The proposed iterative algorithm efficiently solves the CTWSVM optimization problems.

Conclusions:

• The capped L1-norm twin support vector machine (CTWSVM) offers a robust solution for binary classification with outliers.
• CTWSVM retains the benefits of TWSVM while significantly improving resilience to noisy data.
• The developed algorithm and theoretical analysis support the practical application of CTWSVM.