Multiobjective Support Vector Machines: Handling Class Imbalance With Pareto Optimality.

Area of Science:

• Machine Learning
• Optimization Theory

Background:

• Support Vector Machines (SVMs) involve optimizing margin maximization and regularization for both positive and negative classes.
• Appropriate weightage of these objectives is crucial, particularly for imbalanced datasets, often requiring extensive parameter tuning.
• Current methods necessitate costly parameter tuning to find optimal relative weights for SVM objectives.

Purpose of the Study:

• To propose a novel multiobjective optimization framework for training SVMs.
• To overcome the limitations of traditional parameter tuning in SVMs, especially for imbalanced data.
• To introduce the radial boundary intersection method for SVM training.

Main Methods:

• Training Support Vector Machines (SVMs) within a multiobjective optimization framework.
• Utilizing the radial boundary intersection method for optimizing SVM objectives.
• Applying the method to both two-class and multiclass datasets.

Main Results:

• The radial boundary intersection scheme effectively trains SVMs by considering multiple objectives simultaneously.
• Experimental results demonstrate the scheme's effectiveness in finding optimal tradeoffs among SVM objectives.
• The proposed method offers an alternative to traditional, costly parameter-tuning schemes.

Conclusions:

• The radial boundary intersection framework provides an efficient approach to SVM training.
• This method successfully addresses the challenge of balancing multiple SVM objectives, particularly for imbalanced datasets.
• The proposed scheme offers a more effective alternative to conventional parameter tuning for SVMs.