Wasserstein Random Forests and Applications in Heterogeneous Treatment Effects
schedule le lundi 18 mai 2020 de 17h00 à 18h00
Organisé par : F. Bechtold, W. Da Silva , A. Fermanian, S. Has, Y. Yu
Intervenant : Qiming Du (LPSM)
Lieu : online at https://bigbluebutton.math.upmc.fr/b/ade-phf-9dg
Sujet : Wasserstein Random Forests and Applications in Heterogeneous Treatment Effects
We propose a new family of Random Forests that can be used to estimate conditional distributions and we highlight its applications on the estimation of Heterogeneous Treatment Effects. We first provide a brief introduction on Breiman's original RF. Then, we present a reformulation using some simple equivalence between empirical variances and Wasserstein distance of empirical measures, which heuristically indicates that Breiman’s splitting criterion can be used to obtain conditional distribution estimates in both univariate and multivariate cases. Additionally, respecting the samephilosophy of Breiman’s construction, we propose some variant splitting criteria that are more dedicated to the conditional distribution estimation problems. Some preliminary theoretical connections will be revealed, along with various numerical experiments. We expect that the same methodology may also inspire more general supervised learning algorithms such as Boosting and Aggregation, to make them suitable for conditional distribution estimation.