A0660
Title: Additive tree flows for density estimation and two-sample comparison
Authors: Naoki Awaya - Stanford University (United States) [presenting]
Li Ma - Duke University (United States)
Abstract: A new nonparametric method is proposed for two fundamental unsupervised learning tasks: density estimation and two-sample comparison, which are known to be challenging in high-dimensional settings. Motivated by the recent success in normalizing flow methods in the machine learning community, a new class of flow models consisting of "trees", i.e., conditional density functions is introduced, defined on recursive partition structures. The novelty of the proposed method is the introduction of a new efficient sequential algorithm that works like the boosting algorithm typically used for supervised learning. As in the classical boosting algorithm, the proposed algorithm repeatedly transforms the observations ("residuals") and fits a new density function to the residuals. It is shown that the empirical performance of our proposed method is competitive with the deep neural network methods, but the computational cost is drastically improved. Its application is also presented to biological data such as microbiome data.
