EXPERIMENTS WITH SOFT FUSION METHODS WHEN BAGGING k-NN CLASSIFIERS

摘要

We aim to investigate the performance of bagged kNN classifiers when fused using simple fusion methods. Experiments are performed under varying training set sizes including the very small case. Also, we manipulate the feature set size to see if it improves the bagging performance. We are interested in finding which methods of aggregation, in a bagging kNN scenario, would yield a superior performance. The results over different training set sizes show MProduct and Sum to be yield improved bagging performance. However, at very small sample size bagging is successful only when Vote is used. Reducing the feature set leads to improved bagging performance for some of the data sets.