Document classification on neural networks using only positive examples

The training of feedforward neural networks for document filtering was discussed with respect to availability of positive information. A filter was developed for the search of the required document in a corpus of documents. A three-level feedforward neural network with a `bottleneck' was employed under standard back-propagation for learning the identity function. The term-frequency-inverse-document-frequency (tf-idf) and Hadamard product representations were employed for the implementation of the bottleneck filter.