In order to solve the two key problems of the short text classification, very sparse features and strong context dependency, a new method based on latent Dirichlet allocation was proposed. The generated topics not only discriminate contexts of common words and decrease their weights, but also reduce sparsity by connecting distinguishing words and increase their weights. In addition, a short text dataset was constructed by crawling titles of Netease pages. Experiments were done by classifying these short titles using K-nearest neighbors. The proposed method outperforms vector space model and topic-based similarity.

In order to solve the problem of reduction anomaly in the existing probabilistic rough set models, non-parameterized and parameterized maximum decision entropy measures for attribute reduction were proposed by using the concept of maximum confidence of uncertain object. The monotonicity of the parameterized maximum decision entropy was explained and the relationship between its attribute reduction and other ones was analyzed. The definitions for core and relatively dispensable attributes in the proposed model were also given. Moreover, non-parameterized and parameterized confidence discernibility matrixes were put forward and the difference of classical discernibility matrix and the proposed ones in charactering the uncertain object were discussed. Finally, a case study was given to show the validity of the proposed model.