Robust Least Squares Twin Support Vector Machine With Adaptive Pinball Loss.

Area of Science:

• Machine Learning
• Computational Intelligence
• Pattern Recognition

Background:

• Least Squares Twin Support Vector Machine (LSTSVM) offers faster training and better generalization than conventional TSVM.
• LSTSVM's reliance on a square loss function makes it susceptible to noise, degrading generalization.
• Existing methods struggle with noise sensitivity and maintaining classification accuracy.

Purpose of the Study:

• To develop a robust LSTSVM (RLS-TWSVM) with improved anti-noise and generalization capabilities.
• To introduce a novel adaptive pinball loss function that combines benefits of pinball and smooth pinball loss functions.
• To ensure the convexity of the proposed adaptive pinball loss function under specific conditions.

Main Methods:

• Proposed a novel adaptive pinball loss function, guaranteeing convexity under certain conditions.
• Developed a Robust LSTSVM (RLS-TWSVM) by replacing the square loss with the adaptive pinball loss.
• Employed the iteratively reweighted least squares (IRLSs) method for optimization and provided computational complexity and convergence analysis.

Main Results:

• The RLS-TWSVM demonstrates insensitivity to noise near the classification boundary.
• The method simultaneously minimizes within-class scatter and misclassification error.
• Experimental results show superior anti-noise and generalization performance compared to ten related methods, with significant accuracy increases on benchmark and MINIST datasets.

Conclusions:

• The proposed adaptive pinball loss function effectively enhances the robustness and generalization of LSTSVM.
• RLS-TWSVM offers a significant improvement over existing methods, particularly in noisy conditions.
• The method provides a valuable contribution to machine learning for classification tasks with noisy data.