Improving the Performance of Sigmoid Kernels in Multiclass SVM Using Optimization Techniques for Agricultural Fertilizer Recommendation System

Abstract

Support Vector Machines (SVM) are advancing rapidly in the field of machine learning due to their enhancing performance in categorization and prediction. But it is also known that the performance of SVM can be affected by different kernel tricks and regularization parameters like Cost and Gamma. The polynomial kernel seems to be more suitable for performing multiclass SVM classification for the dataset used here. In this study, we propose an improved sigmoid kernel SVM classifier by adjusting the cost and gamma parameters with which a better performance can be achieved. The study is conducted for a multiclass soil fertilizer recommendation system for paddy fields. Furthermore, different optimization methods like Genetic Algorithm and Particle Swarm Optimization are used to tune the SVM parameters. Finally, a comparative study on the performance is also done for the different choices of the parameters, pointing out their accuracies.