THE WEKA MULTILAYER PERCEPTRON CLASSIFIER

Abstract

Automatic document classification is a must when dealing with large collection of documents. WEKA, and especially Weka Knowledge Flow Environment, is a state-of-the-art tool for developing classification applications, even with no programming abilities. We continue our WEKA project presented in a previous paper but changing the classification step, now using the Multilayer Perceptron Classifier. The used dataset is one based on documents from the Reuters Corpus and with vector space model representation, the number of features being reduced by using the InformationGain method. The theoretical bases for Multilayer Perceptron neural networks are presented, both for the architecture and for the backpropagation learning algorithm. In order to evaluate the performance of the Multilayer Perceptron Classifier experiments were done, first with the default network architecture. Results are presented and prove valuable, but for a large number of features the performances decrease. In order to improve the obtained results we test different fine-tuned architectures by changing the number of neurons in the hidden layer. Therefore, the Weka Multilayer Perceptron Classifier is a classifier that deserves attention, but mainly when time requirements are not important at all..