Automatic keyword identification by artificial neural networks compared tomanual identification by users of filtering systems

Information filtering (IF) systems usually filter data items by correlating a vector of terms that represent the user profile with similar vectors of terms that represent data items. Terms that represent data items can be det ermined by experts or automatic indexing methods. In this study we employ a n artificial neural network (ANN) as an alternative method for both IF and term selection and compare its effectiveness to that of "traditional" metho ds. In an earlier study we developed and examined the performance of an IF system that employed content-based and stereotypic rule-based filtering met hods in the domain of e-mail messages. In this study, we train a large-scal e ANN-based filter, which uses meaningful terms in the same database as inp ut, and use it to predict the relevance of those messages. Our results reve al that the ANN relevance prediction out-performs the prediction of the IF system. Moreover, we found very low correlation between the terms in the us er profile (explicitly selected by the users) and the positive causal-index (CI) terms of the ANN, which indicate the relative importance of terms in messages. This implies that the users underestimate the importance of some terms, failing to include them in their profiles. This may explain the rath er low prediction accuracy of the IF system. (C) 2001 Elsevier Science Ltd. All rights reserved.