Research Of Text Categorization Algorithm Based On Non-negative Sparse Representation

Internet has accelerated the rapid development of information technology, we can access to large amounts of information easily and fast every day. However, the explosive growth of information also brings troubles for people to access needed information quickly and efficiently. How to organize and deal with the information effectively to meet the response time and query performance, has become the urgent requirements for the rapid development. Text classification is generated and applied to the full-text retrieval, information driven, pattern recognition systems gradually to obtain a long-term development.In recent years, researchers pay more and more attentions to the issue of non-negative sparse representation and apply it to the image representation. Firstly, we utilize both the local and global information of the sample space and construct a sparse probability graph (SPG) whose nonnegative weight coefficients are derived by nonnegative sparse representation algorithm. The weights of SPG naturally reveal the clustering relationship of labeled samples and unlabeled samples, and meanwhile avoid the adjacency selection and parameter setting process in traditional algorithm of the graph construction.We propose spectral clustering methods and label propagation algorithm for semi-supervised learning based on non-negative sparse representation. In a large sample space, the random and manual labeled samples can not be sufficient to reflect the sample distribution fully and accurately. For this, we propose the spectral clustering algorithm based on non-negative sparse represention, which divide all the samples into several clusters and label the samples that are outstanding and close to the cluster center. Then, Enlightened by LNP method, we propose label propagation algorithm based on non-negative sparse representation which propagates the labels of unlabeled samples algorithm converges. Extensive experimental results on face recognition, UCI machine learning and TDT2 text datasets demonstrate that spectral clustering and label propagation algorithm based on non-negative sparse representation outperforms the standard label propagation algorithm in the aspects of the pression, recall and F1 value.