Access Type

Open Access Dissertation

Date of Award

January 2020

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Computer Science

First Advisor

Dongixao Zhu

Abstract

Predictive modeling (a.k.a. supervised learning) is a machine learning paradigm that has enormous important applications for real-world problems. With the recent surge of data in volume and complexity, effectively capturing the information in input features that is relevant to targets is critical to the success of predictive modeling. Tackling this challenge requires different techniques depending on the specific applications. In this thesis, we develop several methods to improve the performance of predictive models.

Specifically, in the case of small $n$, large $p$ problem, we propose two sparsity-inducing regularization methods for multi-class logistic regression and finite mixture of linear regression, respectively. Those regularization can not only improve predictive performance, but also greatly enhance model interpretability. As an important application of predictive modeling, in healthcare informatics, we develop two DNN models for risk prediction. We also study the learning property of logistic and softmax losses for deep neural networks; we derive a system of equations that quantitatively depicts the property of the decision boundary. Based on this property, an improved version of logistic loss is proposed. Finally, to improve the training of DNNs with stochastic gradient descent, we develop a regularization method of feature alignment via maximum mean discrepancy.

Share

COinS