Contents I. Introduction
Text categorization, which involves assigning one or more predefined categories to a free text according to its content, has turned out to be one of the very important and basic components in text information management. It has been studied for several years, and a number of efficient machine learning approaches have been utilized, such as Bayesian classifiers [1], nearest neighbor classifiers [2], decision trees [1], rule learning [3], support vector machines (SVM) [4], ensemble learning methods [5], neural networks [6], and so on. In machine learning, incomplete data is a big problem. There are many possibilities that can cause the training data to be incomplete, such as mislabeling, biases, omissions, non-sufficiency, imbalance, noise, outliers, etc. As one of the typical machine learning tasks, text categorization consistently suffers from an incomplete training data problem. Among all cases of the data-incomplete circumstance, outlier problem may be the most important factor for text categorization task. An outlier is a pattern that was either mislabeled in the training data, or inherently ambiguous and hard to recognize [7]. In text categorization task, there usually exist a lot of outliers in the training data, for example, non-sense documents, documents mislabeled or lying on the border between different categories, and documents that are out of the defined categories, etc. Therefore, text categorization must address the outlier problem to reach a high performance. However, traditional machine learning approaches mentioned above for text categorization did not take into account the outlier problem in some sense.
