Study on the Optimization of the Bandwidth of the Kernel Function for Kernel Discrimination
Yong Bai · 2009
Kernel density estimation can be used to estimate the conditional class probability density functions and the key problem is to choose the bandwidth of the kernel.In this study,a method that the bandwidth is optimized through maximizing the area under the Receiver Operating Characteristic curve was proposed.Two classification systems called BKD and CKD,which are applicable to binary features and continuous features,respectively,were developed.The prediction accuracies for UCI Promoter dataset and UCI Diabetics dataset by the two classification systems are close to the best literature reported results,showing that application of our proposed method for the optimization of the bandwidth to kernel discrimination has a good prediction ability.