Improve the Efficiency of Support Vector Machine Classifier with Fractional Gradient Descent

Abstract

This study aims to improve the efficiency of the SVM classifier using the fractional descent gradient algorithm, so that the speed during the data training process increases. There are seven numerical datasets (iris, abalone, wine quality, pima Indian diabetes, ionosphere, wheat seeds, sonar) used to test the performance of the SVM classifier with fractional gradient descent which has been optimized with fractional order Caputo Type. The seven datasets have a binary class and a balanced class for the number of labels from one class to another. The problem of minimizing the objective function of the SVM classifier is the problem of infinite minimization which is a convex function that can be solved with Fractional Gradient Descent. Fractional order makes it easier to determine the amount of learning rate used to update weights so as to increase computation time. In this study SVM classifier optimization using a fractional order derivative of the Caputo type, so that the convergence speed is much faster than using a sequence of integers. The test results show that the SVM Classifier with fractional gradient descent, it reaches a convergence point of approximately 60% smaller than traditional SVM gradient descent. SVM classifier with Fractional Gradient Descent has the highest convergence speed in the seven datasets, with a small learning speed in the process of reaching the point of convergence of computing time increases. For the future study we want to use fractional gradient descent SVM for unbalanced class case and sparse dataset that many of the values are zero for text classification.