Robustness of the one-class Peeling method to the Gaussian Kernel Bandwidth

The One Class Peeling (OCP) method is an outlier detection method for Phase I analysis of large multivariate samples. The two stage OCP method is based on support vector data description and employs the Gaussian kernel function to create sequential flexible boundaries to peel away observations to estimate the center of a data set. Subsequently the distance from that estimated center is used as the monitoring statistic. In this work we study the effect of the bandwidth parameter for the Gaussian Kernel on the performance of the OCP method. We find that the OCP method is robust to the value of the bandwidth parameter in terms of in-control and out-of-control simulation performance. We verify these results on example data. The result is a recommendation for a small value of the bandwidth value, providing a computational advantage to the OCP method.