Cross Validation Through Two-Dimensional Solution Surface for Cost-Sensitive SVM.

Area of Science:

• Machine Learning
• Computational Statistics

Background:

• Model selection is crucial for cost-sensitive Support Vector Machines (CS-SVM).
• Traditional methods struggle with CS-SVM's dual regularization parameters, hindering global minimum cross-validation (CV) error computation.
• Existing approaches like grid search are computationally expensive and may not find optimal solutions.

Purpose of the Study:

• To develop an efficient and accurate method for determining the global minimum CV error in CS-SVM.
• To address the challenges posed by CS-SVM's two regularization parameters in model selection.
• To improve the generalization ability and reduce the running time of CS-SVM.

Main Methods:

• Proposed a novel cross-validation approach named CV-SES (solution and error surfaces based CV).
• Developed a bi-parameter space partition algorithm to compute a two-dimensional solution surface for CS-SVM.
• Generated two-dimensional validation error surfaces for each CV fold and superposed them to create a comprehensive CV error surface.

Main Results:

• CV-SES successfully computed the global minimum CV error for CS-SVM.
• Experimental results demonstrated superior generalization ability compared to grid search and other hybrid methods.
• The proposed CV-SES method exhibited significantly reduced running times across multiple datasets.

Conclusions:

• CV-SES provides an effective solution for model selection in CS-SVM, overcoming limitations of existing methods.
• The approach enhances predictive performance and computational efficiency for cost-sensitive and imbalanced learning tasks.
• CV-SES represents a significant advancement in optimizing CS-SVM models.