A0525
Title: Variable importance measure for generalized random forest
Authors: Tomoshige Nakamura - Juntendo University (Japan) [presenting]
Abstract: The extension of variable importance for random forests to Generalized Random Forests (GRFs) is discussed. GRFs are a method for estimating functional parameters defined as the solution of local estimation equations using random forests. However, unlike the conventional random forest case, the ground truth of the parameters cannot be observed, making it impossible to directly compute the Mean Decrease Accuracy (MDA) from the data. Therefore, an Approximate MDA, which approximates the MDA defined by the functional parameters using a score function, is proposed, and based on this, new Permutation MDA and Noise-up MDA are introduced. As an application, the problem of estimating conditional treatment effects is addressed, and the effectiveness of the proposed methods is demonstrated.
