Text Categorization with a Robust Minimax Approach

摘要

This paper explores the use of a new Robust Minimax machine for automated categorization of texts. This Minimax machine minimizes the worst-case (maximum) probability of misclassification of future data points. An important feature of this method considers data in a global fashion, while SVM ignores this kind of global information. We use feature selection based on information gain and other methods to reduce the feature space and improve performance. In this paper, a heuristic one-against-rest strategies for multiclass categorization problem has been proposed by using probability output of the machine and we analyze the particular properties of learning with text data and identify why this approach that use global information is appropriate for this task. Empirical results show that our approach is competitive with the other methods: generally yield high prediction accuracy.