An improved co-training text categorization algorithm based on diversity measures

Co-training algorithm is constrained by its assumption that the features can be split into two compatible and independent subsets. However, the assumption is usually violated in real-world application, especially for independence. We discover its real purpose is to find two classifiers with certain accuracy and sufficient diversity to co-train. First, multi-views are created using different term evaluation functions. Second, instead of directly computing the independence between two sub-views, this paper evaluates the independence between two classifiers, trained on them, by using diversity measures indirectly. Thus a pair of classifiers with certain accuracy and greater diversity is selected. The experimental results show two improved algorithms named TV-SC and TV-DC are both outperform another co-training algorithm named Co-Rnd based on random splitting method, and TV-DC outperforms TV-SC.