An efficient kernel-based nonlinear regression method for two-class classification

In KNRM (kernel-based nonlinear regression model) classifying for one test sample depends on all the kernel functions between each training sample and the test sample. As a result, the classification efficient is enslaved to the size of training set. In this paper KNRM is viewed as a ridge regression model with discrete outputs. Let it be supposed that in feature space discriminant vector can be approximated by some linear combination of a part of training samples (called nodes), then a simple and reasonable algorithm for selecting nodes is developed. Based on the algorithm, DKNRM (derived kernel-based nonlinear regression model) is proposed. When DKNRM classifies one test sample, only kernel functions between each node and the test sample should be calculated. Accordingly, it can be expected that DKNRM will perform well with superiority in classification efficiency. Experimental results on benchmarks show that right classification rates from DKNRM are comparative to naive KNRM, while DKNRM is superior to naive KNRM in classification efficiency.