CMStatistics 2022: Start Registration
View Submission - CMStatistics
B0700
Title: Design-based ensemble learning for individual prediction in finite populations Authors:  Li-Chun Zhang - University of Southampton (United Kingdom) [presenting]
Abstract: One can distinguish between ensemble methods which use a single base learning algorithm to produce homogeneous base predictors (or learners), such as bagging, and those using heterogeneous (component) predictors of different types. The two basic combination methods for heterogeneous ensembles are voting and averaging, whereas gating and stacking are examples of methods which combine by learning a meta-learner. We present a design-based approach to ensemble learning by voting or averaging based on the expected cross-validation results, given the sampling design and the sample-splitting design for cross-validation. Valid inference of the uncertainty of ensemble prediction for finite populations is defined and obtained with respect to the known sampling design, regardless if the assumed model that facilitates prediction is correct or not.