A0187
Title: Jackknife empirical likelihood for infinite-order U-statistics with applications to ensemble predictions
Authors: Qing Wang - Wellesley College (United States) [presenting]
Yichuan Zhao - Georgia State University (United States)
Ting Zhang - University of Georgia (United States)
Abstract: Infinite-order U-statistics have abundant practical applications, such as subsampling-based ensemble methods. However, due to the dependence of the degree kn on the sample size n, theories and results under the traditional fixed-degree U-statistic framework cannot be applied directly. In particular, there has yet to be a promising method to estimate the variance of an infinite-order U-statistic, especially when kn is not of a much lower order of n. The jackknife empirical likelihood methodology is extended to infinite-order U-statistics. It is proven that the fundamental framework of jackknife empirical likelihood still holds under some regularity conditions. In the context of subsampling-based ensemble methods, the performance of the proposed methodology is evaluated to construct confidence intervals for ensemble predictions through simulation studies. The proposal yields superior results compared to existing methods across various settings. In addition, it gives a coverage probability that is approaching the nominal level as the number of trees used to build the ensemble increases.
