A0561
Title: Non-asymptotic rates for random forest prediction intervals via Stein's method
Authors: Krishnakumar Balasubramanian - University of California, Davis (United States) [presenting]
Abstract: Non-asymptotic rates for prediction intervals obtained using random forests using Stein's method will be presented. The previous work on viewing random forest predictions as adaptively weighted k-potential nearest neighbours prediction methods are leveraged to do so. This viewpoint is connected with recent advances in Stein's method for obtaining normal approximation bounds for region-stabilizing statistics to obtain our results for random forests. In particular, it is shown that k-potential nearest neighbours satisfy a certain region-stabilization property. Along the way, refined results on normal approximation were also obtained for a general class of region-stabilizing statistics, improving the recent results of other researchers.
