Generating Heavy-Tailed Synthetic Data with Normalizing Flows

Saba Amiri, Eric Nalisnick, Adam Belloum, Sander Klous, Leon Gommans

Keywords: normalizing flows, mixture models, heavy-tailed data, long-tailed data

TL;DR: In this work, we investigate the ability of normalizing flows to perform heavy-tailed data synthesis

Abstract: Heavy-tailed data is commonly encountered but difficult to model well. We experimentally compare the ability of three different normalizing flows to model data with varying tail behavior. The flows are parameterized using base densities with differing tail properties. We report results on both simulations and a real-world data synthesis task.