Title:	Enhancing Similarity Measures with Imperfect Rule-based Background Knowledge
Authors:	Steffens, Timo
Thesis advisor:	Prof. Dr. Volker Sperschneider
Thesis referee:	Prof. Dr. Ute Schmid Prof. Dr. Ralph Bergmann
Abstract:	Classification is a general framework that can be applied tovarious tasks such as object recognition, prediction, diagnosis or learning. There exist at least two different approaches for classification, namely rule-based and similarity-based classification. The two approaches have different strengths and weaknesses. The former requires a domain theory in order to make inferences from the test instance to its class. The latter does not have this requirement and approximates the class of a test instance via its similarity to a set of known instances.In this thesis the above two approaches are integrated in the realm of Case-Based Reasoning (CBR). CBR treats new cases according to their similarity to stored cases. Similarity is calculated by a similarity measure, which is the crucial factor for classification accuracy. In this work, rule-based domain knowledge is incorporated into the similarity measures of CBR in order to increase classification accuracy. Several novel integration methods are introduced, implemented and evaluated. Since knowledge about real world domains is typically imperfect, the approach does not assume that the domain theories are accurate or complete. Rather, a systematic analysis of different knowledge types is presented that shows the effect of imperfect knowledge on classification accuracy. The analysis is conducted partly empirically in artificial and in real world domains, and partly formally.
URL:	https://osnadocs.ub.uni-osnabrueck.de/handle/urn:nbn:de:gbv:700-2006051214
Subject Keywords:	similarity; classification; case-based reasoning
Issue Date:	7-Jun-2006
Submission date:	7-Jun-2006
Type of publication:	Dissertation oder Habilitation [doctoralThesis]
Appears in Collections:	FB08 - E-Dissertationen

Items in osnaDocs repository are protected by copyright, with all rights reserved, unless otherwise indicated. rightsstatements.org