A0247
Title: Adversarial random forests
Authors: Marvin Wright - Leibniz Institute for Prevention Research and Epidemiology - BIPS (Germany) [presenting]
Abstract: Adversarial random forests are presented, which are a provably consistent tree-based machine learning method designed for density estimation and generative modeling. Inspired by generative adversarial networks, our method employs a recursive procedure in which trees gradually learn structural properties of the data through alternating rounds of generation and discrimination. We achieve comparable or superior performance to state-of-the-art probabilistic circuits and deep learning models on various tabular data benchmarks while executing about two orders of magnitude faster on average. We provide an overview of the methodology, show benchmark results for density estimation and generative modeling, and introduce new methods for conditional sampling. The latter allows applications in missing data imputation and explainable AI, e.g., in conditional feature importance and counterfactual explanations.
