Ensemble algorithms and feature selection

Abstract

A popular technique for modelling data is to construct an ensemble of learners and combine them in to a single hypothesis.  This final model can achieve an accuracy that is greater than that of the ensemble members, provided that there is a sufficient level of diversity within these learners.  Measuring and promoting this diversity can be achieved in a variety of ways and typically a trade-off exists between the accuracy and diversity of the ensemble members.  This thesis investigates and develops ensemble techniques for improving this accuracy and diversity, and compares them to other well-known ensemble methods.  These algorithms are shown to successfully promote diversity whilst maintaining the learner accuracy.

An important area of machine learning research is that of feature selection.  Choosing an appropriate subset of the available features with which to represent the data can improve the performance of learning algorithms in terms of accuracy, efficiency and interpretability.  However, this task is non-trivial and can be complicated further through interactions amongst the features, which can result in features only being relevant within a local area of the space.  Through the creation of diverse local models, ensemble methods have the capacity to address these issues and identify feature relevance.  This work develops new methods that utilise these aspects of ensemble algorithms to identify and exploit feature information.

More information

Identifiers

Catalogue record

Export record

ASCII CitationAtomBibTeXData Cite XMLDublin CoreDublin CoreEP3 XMLEndNoteHTML CitationHTML CitationHTML ListJSONMETSMODSMPEG-21 DIDLOpenURL ContextObjectOpenURL ContextObject in SpanRDF+N-TriplesRDF+N3RDF+XMLRIOXX2 XMLReferReference ManagerSimple Metadata