An Adaptive Non-parametric Kernel Method for Classification

Show full item record

Title: An Adaptive Non-parametric Kernel Method for Classification
Author: Kaszycki, Gregory John
Advisors: Dennis Bahler, Chair
Matthias Stallmann, Member
James Lester, Member
David Dickey, Member
Abstract: One statistical method of estimating an underlying distribution from a set of samples is the non-parametric kernel method. This method can be used in classification toestimate the underlying distributions of the various classes. Since it can be shown that there is no perfect shape to a kernel function used to estimate an underlying distribution, several functions have been proposed and none is superior in all cases. This thesis demonstrates that a function can be created that adapts its shape to fitthe properties of the underlying data set. The function adapts its shape based on a pair of parameters and the algorithm uses a hill-climbing algorithm to determine thebest pair of parameters to use for the data set. This method gets consistently better accuracy than existing non-parametric kernel methods and comparable accuracy toother classification techniques. In addition, this method can estimate information about the underlying characteristics of the data, based on the optimal parameters to thekernel function. This information includes the complexity of the underlying concept function and the general noise level in the measured attributes of the examples.
Date: 2000-03-28
Degree: PhD
Discipline: Computer Science

Files in this item

Files Size Format View
etd.pdf 629.9Kb PDF View/Open

This item appears in the following Collection(s)

Show full item record