Comparison of the Feature-Based Combiner to Bagging and RSM using Artificial Data

Abstract

We experimentally compare the feature based combiner to the most common combiner methods of bagging and random subspace method. The experiments are made on different synthetic data sets to find when the FBC outperforms bagging and RSM. Results show that FBC outperforms other combiners when a small number of features exist, especially when the number of combined classifiers is low. As the number of combined classifiers increase, it underperforms other combiners. We mainly find that when number of classes increases some classes may not be well represented in the training set. This is where FBC performs worse than other methods. The problem increases at smaller number of samples. This is an obvious consequence because the possibility of class misrepresentation increases as the size of the training set decreases.