Kernel-Based Text Classification on Statistical Manifold

In the text literature, a variety of useful kernel methods have been developed by many researchers. However, embedding text data, into Euclidean space is the key characteristic of common kernels-based text categorization. In this paper, we focus oil representation text; vectors as points on Riemann manifold and use kernels to integrate discriminative and generative model. And then, we present diffuse kernel based oil Dirichlet Compound Multinomial manifold (DCM manifold) which is a space about Dirichlet Compound Multinomial model combining inverse document frequency and information gain. More specifically, as demonstrated by our experimental results on various real-world text datasets, we show that the kernel based oil this DCM manifold is more desirable than Euclidean space for text categorization. And our kernel method provides much better computational accuracy, than some current state-of-the-art methods.