Abstract

In this thesis we devise a method for performing distribution regression utilizing an extension of mixture density networks. This method will be benchmarked against other state-of-the-art methods (utilizing negative log loss) and distribution-agnostic models of similar complexity. We demonstrate that our proposed method performs at least as well as the state-of-the-art methods in distribution regression. Additionally, we show that adding distribution predictive capabilities does not significantly decrease performance in models. We illustrate how one can utilize this methodology to regress to confidence interval predictions. Finally, we demonstrate a variety of novel and interesting applications of our framework.

Library of Congress Subject Headings

Regression analysis; Distribution (Probability theory); Machine learning; Uncertainty--Mathematical models

Publication Date

5-2019

Document Type

Thesis

Student Type

Graduate

Degree Name

Computer Science (MS)

Department, Program, or Center

Computer Science (GCCIS)

Advisor

Ifeoma Nwogu

Advisor/Committee Member

Rui Li

Advisor/Committee Member

Qi Yu

Campus

RIT – Main Campus

Plan Codes

COMPSCI-MS

Download

Share

COinS