Author

Xuanfu Wu

Date of Award

5-1-2004

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Science

First Advisor

Dr. Zhengxin Chen

Abstract

An ensemble consists of a set of individual trained classifiers whose individual decisions are combined in some way to classify new examples. Many ensemble techniques have been developed with great success. This type of technique generates several classifiers to form an ensemble through meta-learning techniques. At the classification stage, the ensemble combines the results of individual classifiers to make the final decision. However, existing studies typically take a static approach in assembling individual classifiers and have mainly focused on either ensemble generating or ensemble combination. In this paper, after a critical examination of existing approaches, a new concept of dynamic ensemble is proposed. A dynamic ensemble is an advanced classifier that could have dynamic component classifiers and have dynamic configurations. Toward this goal, we have substantially expanded the existing "overproduce and choose" paradigm for ensemble construction. A new algorithm toward dynamic ensembles, namely BAGA, is presented in this paper. BAGA takes a set of decision tree component classifiers as input, and generates a set of candidate ensembles. Rather than goes to the classification stage immediately as does in most existing approaches, BAGA introduces another layer of learning phase to select the optimal subset of classifiers using genetic algorithm to constitute the dynamic ensemble. Experiments results on several datasets showed that BAGA can achieve a better performance than traditional bagging (which uses a fix number of classifiers). In this paper, the effects of bag size, voting function and evaluation functions on the dynamic ensemble construction are investigated. This paper also provides a guideline for applicability of the proposed algorithms, which offers criteria for future work of selecting appropriate configurations for dynamic ensembles. Other aspects for future work in dynamic ensembles are also discussed.

Comments

A Thesis Presented to the Department of Computer Science and the Faculty of the Graduate College University of Nebraska In Partial Fulfillment of the Requirements for the Degree Master of Science University of Nebraska at Omaha Copyright 2004 Xuanfu Wu

Files over 3MB may be slow to open. For best results, right-click and select "save as..."

COinS