An evolutionary approach for discovering effective composite features for text categorization

The study of text categorization has assumed special significance in the Internet era in helping us navigate the ocean of web pages and emails that continue to grow in an unrelenting pace. In many previous works on text classifications, it has been shown that composite features consisting of multiple word tokens like statistical phrases can contribute effectively to the classification task. However finding useful composite features through comprehensive search from the vast number of possibilities is often prohibitive in terms of computing resource requirements. In the past, to make the search feasible, we often limit the search space by imposing some parametric constraints like minimum frequency and/or number of words in the composite feature. In this paper we proposed a new evolutionary approach to find effective composite features for classification, an approach that combines probabilistic feature generation with error-biased sampling We demonstrate the effectiveness of our approach using the Reuters-21578 test collection.