Proceedings of The 2nd International Conference on Innovation in Computer Science and Artificial Intelligence
An Experimental Comparison of Classifier Combining Methods Using Artificial Data
Fuad M. Alkoot
We experimentally compare the performance of four widely used combiner methods with the aim to show when our previously proposed method outperforms existing methods. Recently we have proposed a novel combiner method to detect autism. The proposed method outperformed existing combiner methods. However, we did not identify the reason behind its outstanding performance. We aim at finding when and why this method outperformed existing methods of bagging, boosting and random subspace methods. We repeat the experiments for varying number of classes, training set sizes and number of features using carefully designed artificial data.
Keywords: Classifier combining, bagging, feature based combiner, random subspace method, neural network.