Model Construction with Support Vector Machines and Gaussian Processes through Kernel Search
Fredrik Hjorth Bentsen · NORA - Norwegian Open Research Archives · 2019
In this thesis, we aim at constructing a framework for kernel searching in both Gaussian process and Support Vector Machines.Choosing the right kernels for these two methods often requires expert knowledge and many people who use these methods do not have enough knowledge of making a good choice at first hand.A system which is automating the choice of kernels could be useful for non-experts and this project carries out an experimental study of how the automatic chosen system can be formulated according to the data structure.Based on our system, we have carried out four empirical analyses: two in regression and two in classification.To evaluate the constructed kernels and final models in the empirical cases, we have followed an experimental and innovative approach instead of a traditional approach.The implementation of the kernel search and model evaluation is explained in detail.More concretely, the data sets we have used in regression are time series data and we have focused on finding models which have reasonable extrapolations.In the classification analyses, we have chosen medical data where we want to find out whether a patient has a disease or not.We have trained the classification models to have good accuracy but also to minimize the type II error of not identifying a patient with the disease.