Authors

Jiaju Yan

Type

Text

Type

Dissertation

Advisor

Kuan, Pei Fen | Zhu, Wei | Wu, Song | Xiao, Keli.

Date

2017-05-01

Keywords

AUC, Imbalanced Classification, Model Ensemble, Random Forest, ROC, Tree Based Method | Statistics

Department

Department of Applied Mathematics and Statistics

Language

en_US

Source

This work is sponsored by the Stony Brook University Graduate School in compliance with the requirements for completion of degree.

Identifier

http://hdl.handle.net/11401/77355

Publisher

The Graduate School, Stony Brook University: Stony Brook, NY.

Format

application/pdf

Abstract

The imbalanced class problem in classification is highly relevant in many realistic scenarios such as the detection of a rare condition. One solution is to design specific algorithms incorporating the unbalanced classes in the training process of a classifier. In this dissertation, we propose a novel multi-class classification tree based on the area under the ROC curve (AUC) to resolve the imbalanced classification problem. This tree classifier aims to maximize the sum of AUC for all one versus all classifiers at the node attribute selection stage while balancing the performance of sensitivity and specificity of all one versus all classification at the node threshold selection stage. The ROC tree is extended to ROC random forest with suitable modifications. Furthermore, the volume under surface (VUS), the extension of AUC for multi-class classification, is discussed in this dissertation as well and used to measure the performance of classifiers. The simulation results show that this multi-class ROC tree/forest method is superior to the classic CART/random forest on severely imbalanced multi-class classification problems, while the ROC random forest performs equally well as the SMOTE random forest on imbalanced binary classification problems. The application on Boston housing data shows that the ROC random forest can also be used for model ensemble and it performs better than all the base models and other ensemble methods in this application. | 100 pages

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.