To improve scalability of text categorization and reduce over-fitting, it is desirable to reduce the number of words used for categorisiation. Further, it is desirable to achieve such a goal automatically without sacrificing the categorization accuracy. Such techniques are known as automatic feature selection methods. Typically this is done in the way that each word is assigned a weight (using a word scoring metric) and the top scoring words are then used to describe a document collection. There are several word scoring metrics which have been employed in literature. In this paper we present a novel feature selection method called the GU metric. The details of comparative evaluation of all the other methods are given. The results show that the GU metric outperforms some of the other well known feature selection methods.
展开▼