One-class Document Classification

In this book, we address the problem of automated §information retrieval and document classification §using only positive examples.In this book, we show §how a simple feed-forward neural network can be §trained to filter documents under these conditions,§and that this method seems to be superior to modified§methods (modified to use only positive training §examples),such as Rocchio, Nearest Neighbor,§Naive-Bayes and Distance- based Probability §algorithms.A novel experimental finding§is that retrieval is enhanced substantially§in this context by carrying out a certain kind of §uniform transformation (Hadamard) of the §information prior to the training of the network.§We also implemented versions of support vector §machine (SVM) appropriate for one-classclassification§in the context of information retrieval.§Finally we present a system designed to help a user §navigate the Web. The system is built upon neural §network techniques designed to attack the problem of §user modeling using only positive examples.