A Model for Term Selection in Text Categorization Problems

In the last ten years, automatic Text Categorization (TC) has been gaining an increasing interest from the research community, due to the need to organize a massive number of digital documents. Following a machine learning paradigm, this paper presents a model which regards TC as a classification task supported by a wrapper approach and combines the utilization of a Genetic Algorithm (GA) with a filter. First, a filter is used to weigh the relevance of terms in documents. Then, the top-ranked terms are grouped in several nested sets of relatively small size. These sets are explored by a GA which extracts the subset of terms that best categorize documents. Experimental results on the Reuters-21578 dataset state the effectiveness of the proposed model and its competitiveness with the learning approaches proposed in the TC literature.