Boosting is a new learning technique that has become very attractive to many researchers involved in the areas of machine learning and statistical pattern recognition. The general idea of boosting is to develop the classifier team incrementally, adding one classifier at a time. The classifier that joins the ensemble at each step is trained on a data set selectively sampled from the training data set. The sampling distribution starts out uniform, and progresses toward increasing the likelihood of “difficult” data points. Thus, the distribution is updated at each step, increasing the likelihood of the objects misclassified at the previous step.
The authors propose an algorithm for boosting kernel density estimation, and show that boosting kernel classifiers reduces the bias with an overall reduction in error. Based on the main result of the paper, a series of conclusions explaining the good performance of boosting kernel classifiers is derived.
The paper is structured into six sections. The general framework of the nonparametric estimation method of density functions using kernels is briefly presented in the introductory section. Since, in the case of the two-class discrimination problem of equal a priori probabilities, the Bayesian classifier reduces to the difference g(x) of the values of the class density functions, the bias and variance of its estimate near the solution g(x0 )=0 are next evaluated when the same kernel function is used to estimate both density functions. In the third section, BoostKDC, a boosting algorithm for kernel density estimation, is proposed, and it is proven that boosting the kernel classifier reduces the bias, with an overall reduction in error. The overfitting effect in boosting is investigated, and cross-validation is proposed as a possible approach to preventing the overfitting. The analysis performed in the fourth section yields convincing explanations about the effect of bias reduction achieved by boosting when it is used in kernel discrimination. In the final section of the paper, comparisons of boosting with simple kernel methods are discussed and supported by simulations and experimental tests.