A simple and reliable instance selection for fast training support vector machine: Valid Border Recognition

Area of Science:

• Machine Learning
• Computational Statistics

Background:

• Support vector machines (SVMs) face training complexity challenges with large datasets.
• Existing instance selection (IS) methods struggle to balance accuracy and computational efficiency.

Purpose of the Study:

• To develop novel instance selection methods for improving SVM training efficiency.
• To address the limitations of current IS techniques in handling large-scale data.

Main Methods:

• Introduced Valid Border Recognition (VBR) to select critical instances based on heterogeneous neighbors.
• Developed a strengthened version (SVBR) that refines instance selection for improved reliability.
• Incorporated IS into Gaussian kernel matrix reduction to minimize execution time.

Main Results:

• VBR and SVBR demonstrated effectiveness in reducing training and inference times.
• Proposed methods maintained or improved classification accuracy compared to existing approaches.
• Experimental validation on benchmark and synthetic datasets confirmed the efficacy of VBR and SVBR.

Conclusions:

• VBR and SVBR offer a viable solution for efficient SVM training on large datasets.
• The proposed methods successfully balance accuracy and computational efficiency in instance selection.