Bagging KNN Classifiers using Different Expert Fusion Strategies
Amer. J. AlBaghdadi, Fuad M. Alkoot · 2005
An experimental evaluation of Bagging K-nearest neighbor classifiers (KNN) is performed. The goal is to investigate whether varying soft methods of aggregation would yield better results than Sum and Vote. We evaluate the performance of Sum, Product, MProduct, Minimum, Maximum, Median and Vote under varying parameters. The results over different training set sizes show minor improvement due to combining using Sum and MProduct. At very small sample size no improvement is achieved from bagging KNN classifiers. While Minimum and Maximum do not improve at almost any training set size, Vote and Median showed an improvement when larger training set sizes were tested. Reducing the number of features at large training set size improved the performance of the leading fusion strategies.