Matching SVM kernel's suitability to data characteristics using tree by fuzzy C-means clustering

Over the last decade kernel based learning algorithms known as Support Vector Machines (SVMs) have become an attractive tool to solve pattern recognition problems. Choosing an appropriate kernel still is a trial and error approach for SVM however. This research provides some insights into the data characteristics that suit particular kernels. Our approach consists of four main stages. First, the performance of six kernels is examined across a collection of 33 classification problems from the machine learning literature. Secondly, a collection of statistics that describe each of the 33 problems in terms of data complexity is collected. After that, fuzzy C-means (FCM) is used to cluster, and construct a decision tree is used to generate the rules of the 33 problems based on these measurea of complexity. Each cluster represents a group of classification problems with similar data characteristics. The performance of each kernel within each cluster and the rules among the tree is then examined in the final stage to provide both quantitative and qualitative insights into which kernels perform best on certain problem types.