Publication Date

Summer 6-28-2022

Abstract

In this dissertation, I propose new approaches to multi-task learning, inspired by statistical model diagnostics and semiparametric and additive modeling. The newly designed additive multi-task model framework allows for flexible estimation of multi-task parametric and nonparametric effects by using an extension of the backfitting algorithm. Further, I propose new methods for statistical task diagnostics, which allow for the identification and remedy of outlier tasks, based on task-specific performance metrics and their empirical distributions. I perform a deep examination of the well-established multi-task kernel method and achieve theoretical and experimental contributions. Lastly, I propose a two-step modeling approach to multi-task modeling, where the tasks are modeled differently according to their belongingness to different clusters based on the selected performance criteria. The newly proposed frameworks are examined on a well-known real-world multi-task benchmark dataset and show significant improvement over other modern multi-task learning methods.

Degree Name

Statistics

Level of Degree

Doctoral

Department Name

Mathematics & Statistics

First Committee Member (Chair)

Guoyi Zhang

Second Committee Member

Yan Lu

Third Committee Member

Fletcher Christensen

Fourth Committee Member

Manel Martínez-Ramón

Language

English

Keywords

multi-task learning, task diagnostics, additive models, outlier tasks, backfitting algorithm, task clustering

Document Type

Dissertation

Download

Share

COinS