Greedy Term Selection for Document Classification with Given Minimal Precision

In this paper we present a new term selection technique for document clustering. Classifiers are trained to recognize documents of one single topic. The technique is designed to maintain a predefined minimal precision, while maximizing the recall of the classification. The most useful terms are collected in a greedy way. The minimal precision is ensured due to a minimal document score for selection, which is calculated based on a separate training set after the term set optimization.