Semi-supervised Co-training Classification Research And Its Application

Generally speaking, only with a great number of labeled data can the conventional supervised classification algorithm make a good model and thus get a good classifying. However, it cannot be satisfied in the world, which makes the semi-supervised classification method significant to discover the information contained in the unlabeled data.This paper studies the Co-training algorithm, one of the most important semi-supervised classification methods. In the composite structure, two traditional classifiers are improved, an arbitration classifier is added to make up for deficiencies in the first two classifiers, and the corresponding method of classifier selection is put forward. Furthermore, in the field of dataset, a new processing method is proposed. It measures for the classifier organization, and constructs a most favorable training set, in which the information loss caused by the deleted point will be made up from the unlabeled data by the semi-supervised classification method.In the process of semi-supervised classifying, the effect will be affected by the error flags. This paper introduces an improved nearest rule method to examine every new labeled point, in which the error flags are unconcluded, the robustness of the model is improved and the security of selected unlabeled data is guaranteed. A part of the unlabeled data are deleted in the process, but the number will be quite limited, and the SMOTE algorithm will be adopted to add the amount of information and make up the information loss caused in g guaranteeing the security.The experimental results show that this algorithm has better performance than the traditional one.