\contentsline {chapter}{\numberline {1}Introduction}{1}
\contentsline {section}{\numberline {1.1}Statement of Thesis}{4}
\contentsline {section}{\numberline {1.2}Contributions of this Thesis}{4}
\contentsline {section}{\numberline {1.3}Structure of this Thesis}{4}
\contentsline {chapter}{\numberline {2}Background and Related Work}{5}
\contentsline {section}{\numberline {2.1}Prototype Learning for Nearest Neighbor Classifiers}{6}
\contentsline {subsection}{\numberline {2.1.1}Instance Selection}{6}
\contentsline {subsubsection}{Condensed Nearest Neighbor (CNN)}{6}
\contentsline {subsubsection}{Reduced Nearest Neighbor (RNN)}{8}
\contentsline {subsubsection}{Minimal Consistent Set (MCS)}{8}
\contentsline {subsection}{\numberline {2.1.2}Instance Abstraction}{9}
\contentsline {subsubsection}{Chang}{9}
\contentsline {section}{\numberline {2.2}Evaluation of Prototype Learning Schemes}{10}
\contentsline {subsection}{\numberline {2.2.1}Storage Reduction}{10}
\contentsline {subsection}{\numberline {2.2.2}Speed Increase}{10}
\contentsline {subsection}{\numberline {2.2.3}Generalization Accuracy}{10}
\contentsline {subsection}{\numberline {2.2.4}Noise Tolerance}{10}
\contentsline {subsection}{\numberline {2.2.5}Probability of Detection and False Alarm}{10}
\contentsline {chapter}{\numberline {3}CLIFF: Tool for Instance Selection}{12}
\contentsline {section}{\numberline {3.1}CLIFF}{13}
\contentsline {section}{\numberline {3.2}CLIFF: A Simple Example}{15}
\contentsline {section}{\numberline {3.3}CLIFF: Time Complexity}{16}
\contentsline {chapter}{\numberline {4}CLIFF Assessment}{17}
\contentsline {section}{\numberline {4.1}Data and Preprocessing Tools}{17}
\contentsline {subsection}{\numberline {4.1.1}Data Set Characteristics}{17}
\contentsline {subsection}{\numberline {4.1.2}Pre-processing tools for Dimensionality Reduction}{18}
\contentsline {subsubsection}{FastMap}{18}
\contentsline {subsubsection}{Feature Subset Selection (FSS)}{19}
\contentsline {section}{\numberline {4.2}CLIFF Assessment on Standard Data Sets}{22}
\contentsline {subsection}{\numberline {4.2.1}Data}{22}
\contentsline {subsection}{\numberline {4.2.2}Experimental Method}{22}
\contentsline {subsection}{\numberline {4.2.3}Experiment ####1: Is CLIFF viable as a Prototype Learning Scheme for NNC?}{23}
\contentsline {subsection}{\numberline {4.2.4}Experiment ####2: How well does CLIFF handle the presence of noise?}{23}
\contentsline {chapter}{\numberline {5}Case Study: Solving the Problem of Brittleness in Forensic Models}{31}
\contentsline {section}{\numberline {5.1}Introduction}{31}
\contentsline {section}{\numberline {5.2}Visualization of Brittleness}{33}
\contentsline {section}{\numberline {5.3}Glass Forensic Models}{33}
\contentsline {subsection}{\numberline {5.3.1}Seheult 1978}{34}
\contentsline {subsection}{\numberline {5.3.2}Grove 1980}{35}
\contentsline {subsection}{\numberline {5.3.3}Evett 1995}{36}
\contentsline {subsection}{\numberline {5.3.4}Walsh 1996}{37}
\contentsline {section}{\numberline {5.4}Visualization of Brittleness in Models}{38}
\contentsline {section}{\numberline {5.5}Introduction}{40}
\contentsline {section}{\numberline {5.6}Dimensionality Reduction}{42}
\contentsline {subsection}{\numberline {5.6.1}Principal Component Analysis}{42}
\contentsline {section}{\numberline {5.7}Clustering}{45}
\contentsline {section}{\numberline {5.8}Classification with KNN}{45}
\contentsline {section}{\numberline {5.9}The Brittleness Measure}{47}
\contentsline {section}{\numberline {5.10}Data Set and Experimental Method}{48}
\contentsline {section}{\numberline {5.11}Experiment 1: KNN as a forensic model?}{49}
\contentsline {subsection}{\numberline {5.11.1}Results from Experiment 1}{49}
\contentsline {section}{\numberline {5.12}Experiment 2: Can brittleness be reduced?}{50}
\contentsline {subsection}{\numberline {5.12.1}Results from Experiment 2}{52}
\contentsline {chapter}{\numberline {6}Conclusion}{56}