A new approach to classifier ensemble learning based on clustering

Abstract

The problem of combining multiple classifiers, referred to as a classifier ensemble, is one sub-domain of machine learning which has received significant attention in recent years. A classifier ensemble is an integration of classification models, referred to as base classifiers, whose individual decisions are combined in order to obtain a final prediction. The aim of using a classifier ensemble is to provide an overall level of performance which is superior to the performance of any of the single base classifiers. This Thesis studies the problem of constructing a classifier ensemble from different perspectives with the aim of improving the overall level of performance.

Two novel ensemble techniques were introduced and evaluated within this study. The first approach, referred to as Classification by Cluster Analysis, was proposed as an alternative solution to the Stacking technique. The new method applies a clustering technique for the purpose of combining base classifier outputs. This approach offers benefits with reduced classification time compared with existing ensemble methods. In addition, it outperformed other ensemble methods in terms of classification accuracy. As an extension to the concept the method was adapted to incorporate semi-supervised learning which is subsequently considered as a new research direction within the domain of ensemble learning.

The second method, referred to as Cluster-Based Classifier Ensemble, was proposed as an alternative to the Nearest Neighbour classifier ensemble. It applies a clustering technique for the purpose of generating base classifiers. A new combining function was proposed to be applied with the method as an alternative to the conventional majority voting technique. The new approach outperforms existing ensemble methods in terms of accuracy and efficiency. Both methods were evaluated in an activity recognition

Both methods were evaluated in an activity recognition problem considered within the work as a case study. The effectiveness of the two methods was further supported by the findings from an experimental evaluation with a real world data set.

Thesis is embargoed until 31st October 2014

Date of Award	Oct 2012
Original language	English
Supervisor	Yaxin Bi (Supervisor), Shengli Wu (Supervisor) & Christopher Nugent (Supervisor)