Prior-knowledge based Green's kernel for support vector regression.

摘要

This thesis presents a novel prior knowledge based Green's kernel for support vector regression (SVR) and provides an empirical investigation of SVM's (support vector machines) ability to model complex real world problems using a real dataset. After reviewing the theoretical background such as theory of SVM, the correspondence between kernels functions used in SVM and regularization operators used in regularization networks as well as the use of Green's function of their corresponding regularization operators to construct kernel functions for SVM, a mathematical framework is presented to obtain the domain knowledge about the magnitude of the Fourier transform of the function to be predicted and design a prior knowledge based Green's kernel that exhibits optimal regularization properties by using the concept of matched filters. The matched filter behavior of the proposed kernel function provides the optimal regularization and also makes it suitable for signals corrupted with noise that includes many real world systems. Several experiments, mostly using benchmark datasets ranging from simple regression models to non-linear and high dimensional chaotic time series, have been conducted in order to compare the performance of the proposed technique with the results already published in the literature for other existing support vector kernels over a variety of settings including different noise levels, noise models, loss functions and SVM variations. The proposed kernel function improves the best known results by 18.6% and 24.4% on a benchmark dataset for two different experimental settings.