A distance-based kernel for classification via Support Vector Machines.

Area of Science:

• Machine Learning
• Data Mining
• Computational Statistics

Background:

• Support Vector Machines (SVMs) are established supervised learning algorithms for classification.
• Traditional SVM methods often use fixed training and testing data splits.
• Existing kernels may not optimally handle diverse feature types in classification.

Purpose of the Study:

• To develop an innovative SVM classification approach using iterative random subset selection.
• To introduce a novel distance-based kernel tailored for binary and multi-class classification.
• To enhance classification accuracy and population inference in machine learning.

Main Methods:

• Employing iterative training with randomly selected data subsets to identify representative samples.
• Designing a new distance-based kernel utilizing a similarity matrix for binary and multi-class features.
• Conducting computational experiments on diverse, publicly available datasets.

Main Results:

• The proposed iterative subset selection method significantly improves classification accuracy over traditional approaches.
• The novel distance-based kernel demonstrates superior performance compared to existing kernels.
• The method shows effectiveness across datasets of varying sizes and complexities.

Conclusions:

• The developed SVM classification method and distance-based kernel offer significant improvements in accuracy.
• This approach enhances the reliability of inferences about the population from data.
• The findings have broad implications for machine learning and data analysis classification problems.