SINGLE TO MULTIPLE KERNEL LEARNING WITH FOUR POPULAR SVM KERNELS (SURVEY)

Yassein Eltayeb Mohamed Idris, Li Jun

Abstract: Machine learning applications and pattern recognition have gained great attention recently because of the variety of applications depend on machine learning techniques, these techniques could make many processes easier and also reduce the amount of human interference (more automation). This paper research four of the most popular kernels used with Support Vector Machines (SVM) for Classification purposes. This survey uses Linear, Polynomial, Gaussian and Sigmoid kernels, each in a single form and all together as un-weighted sum of kernels as form of Multi-Kernel Learning (MKL), with eleven datasets, these data sets are benchmark datasets with different types of features and different number of classes, so some will be used with Two-Classes Classification (Binary Classification) and some with Multi-Class Classification. Shogun machine learning Toolbox is used with Python programming language to perform the classification and also to handle the pre-classification operations like Feature Scaling (Normalization).The Cross Validation technique is used to find the best performance Out of the suggested different kernels methods .To compare the final results two performance measuring techniques are used; classification accuracy and Area Under Receiver Operating Characteristic (ROC). General basics of SVM and used Kernels with classification parameters are given through the first part of the paper, then experimental details are explained in steps and after those steps, experimental results from the steps are given with final histograms represent the differences of the outputs accuracies and the areas under ROC curves (AUC). Finally best methods obtained are applied on remote sensing data sets and the results are compared to a state of art work published in the field using the same set of data

Keywords: Machine Learning, Classification, SVM, MKL, Cross Validation and ROC

DOI: https://doi.org/10.15623/ijret.2016.0503066