A0248
Title: Wasserstein generative regression
Authors: Jian Huang - The Hong Kong Polytechnic University (China) [presenting]
Abstract: A Wasserstein generative regression (WGR) approach is proposed to learn a general regression function nonparametrically using deep neural networks. WGR is based on an objective function constructed by regularizing the usual least squares loss with the Wasserstein distance between the distribution of the regression function and the data distribution. WGR learns a general regression function that can also serve as a generator for sampling from the conditional distribution of the response given the predictor. This is different from the usual regression methods that only learn the conditional mean or conditional quantile functions of the response given the predictor. Another attractive feature of WGR is that it can easily handle high-dimensional responses and predictors. Some preliminary results on the consistency and non-asymptotic error bounds of WGR under appropriate conditions are presented. Extensive numerical experiments are also conducted to demonstrate the advantages of WGR.
