A0868
Title: Stability of random forests and random-forest powered prediction
Authors: Yan Wang - Wayne State University (United States) [presenting]
Abstract: The stability property of random forests that hold even for its greedy version, which is practically implemented in popular packages, is first presented. It is then shown that, based on its stability, the random forest algorithm can be conveniently used to construct prediction intervals with guaranteed marginal coverage under mild conditions and without additional computation. Moreover, it turns out that the stability property can also be taken advantage of in the settings of active and semi-supervised learning using random forests, the guiding principle of which is bias correction rather than variance reduction, as in many existing algorithms. The general methodology presented here is anticipated to be applicable across a wide range of scenarios in medicine and engineering.
