An Adaptive Meta-classifier for Text Documents

In this paper we investigated a way to create a new adaptive meta classifier for classifying text documents in order to increase the classification accuracy. During the first processing phase (pre classification) the meta-classifier uses a non-adaptive selector. The role of this selector is to implement a data transformation from a large space representation of the documents (in our case vectors having 1309 words) into a much smaller space representation, based on the input data categories (in our case there are 16 categories). This transposition method is based on the categories in which the input data might be classified (Reuter's classification) and it is using SVM and Bayes type classification algorithms. In the second phase (classification) we use a feed-forward neural network based on the back-propagation learning method. We chose a neural network architecture which contains one hidden layer of units with sigmoid activation function, where each unit from each layer is connected with all units of the previous layer. The experimental results have showed that using this adaptive algorithm, classification accuracy can be significantly improved. For Reuters2000 text documents we obtained classification accuracy up to 99.74%.