Influence of Metric on Classification Error of Distance-Based Classifiers

0436236 - ÚI 2015 CZ eng V - Výzkumná zpráva
Jiřina, Marcel
Influence of Metric on Classification Error of Distance-Based Classifiers.
Prague: ICS AS CR, 2014. 25 s. Technical Report, V-1211.
Institucionální podpora: RVO:67985807
Klíčová slova: multidimensional data * classifier * distance * metrics * Hassanat metrics * k-NN * IINC
Kód oboru RIV: BB - Aplikovaná statistika, operační výzkum

Five types of classifiers that use sample distances for class estimation of an unknown sample was tested. Each classifier was tested with fifteen different metrics on 24 classification tasks from the UCI Machine Learning Repository. The metrics were compared and the best of them was found for each classifier. Surprisingly, the best metrics for all five types of classifiers is the Hassanat metrics. Classifiers were also compared and ranked according to their classification ability. Wilcoxon Test and Friedman Aligned test were used for statistical evaluation.
Trvalý link: http://hdl.handle.net/11104/0240014