| According to the development trend, Internet will be the main source from which people get information. However, large number of information accumulates in the Internet because of the rapid increasing of it. But the information is disorder without necessary organized, and it makes people difficult to obtain the useful resource. The information categorization can help people locating the information accurately and rapidly, resolves the problem very well. Because the text is the main information carrier, so that a well text auto-categorization system can organize and manage the information availably, supporting the information extracting effectively. Text classification is the process of assigning predefined category labels to new documents based on the classifier learnt from training examples. In traditional classification, training examples are labeled with the same set of pre-defined category or class labels and labeling is often done manually. That is, the training examples set is composed of labeled positive examples set and negative examples set. In recent year, A number of statistical classification and machine learning techniques has been applied to text categorization, including regression models, nearest neighbor classifiers, decision tree, Bayesian classifiers, support vector machines, rule learning algorithm, relevance feedback, voted classification, neural networks,etc. The main problem with this classic approach is that a large number of labeled training examples are needed for accurate learning. Since labeling typically done manually, it is labor intensive and time consuming. Collecting negative training examples is especially delicate and arduous because (1) negative training examples must uniformly represent the universal set excluding the positive class, and (2) manually collected negative training examples could be biased because of human's unintentional prejudice, which could be detrimental to classification accuracy. In recent years, researchers investigated the idea of using a small labeled set of positive class and a large unlabeled set to help learning. This reduces the manual labeling effort. In many real-world learning problems, it is common for the study on PU problem. In general, unlabeled data and positive data are easy collect, and negative data is rare and expensive. For example, consider the automatic diagnosis of disease: unlabeled data are easy to collect (all patients in the database), and positive data are also readily available (the patients who have the disease), but negative data expensive if detection tests for the disease are expensive since all patients in the database can not be assumed to be negative samples if they have never been tested. From this, the problem of PU not only exits the field of text classification, but also has the extensively applied foreground in the other aspects of information retrieval. The thesis studied the PU-oriented text classification. First, we introduced the definition of text classification, the meaning of PU problem study and the correlated work in the field. Second, we discussed the pivotal technology of text classification in detail, such as text expression, feature selection, usual classify model and the parameter for evaluating the performance of a classifier. The principle part of this paper is the study of PU-oriented text classification, analyzed the difference between it and classical text classification methods, explained the framework of the PU-oriented text classifier in detail, and summarized two methods to solve the problem: constructing text classifier only using positive data and constructing text classifier with the frame of two steps. We also did many experiments and the results proved that the former method did not utilize unlabeled data set, so that the effect of the classifier is worse than the second method. During the study of extracting reliable negative data from unlabeled data set, we found that the 1-DNF algorithm defines the positive feature only based on its appearance frequency in positive set and unlabeled set, which results in too many positive features and little negative examples, and this is harmful to the convergence speed and precision of the next step. After analyzed the 1-DNF algorithm, we improved it by defining the positive feature not only using the frequency of occurring in the positive set and unlabeled set, but also restricting the frequency of the feature occurring in the... |
