17 / 2021-06-15 17:57:03

Support Vector Classifier Trained by Gradient Descent

In this study, gradient descent (GD) based support vector classifier (SVC) is proposed for binary classification. There are two structures for the proposed method, i.e., the primal form and the dual form which will be trained using gradient descent. Comprehensive comparisons among GD based SVC, GD based full connected (FC), and traditional training method of SVC are conducted. Experimental results for six datasets demonstrate that the accuracy of GD based primal form SVC is better than FC, and close to traditional train method of SVC. The GD based dual form SVC performs better in small samples. While the proposed SVCs have better generalization ability than FC.