The rapid growth of XML adoption has urged for the need of a proper representation for semi-structured documents, where the document structural information has to be taken into account so as to support more precise document analysis. In this paper, an XML document representation named "structured link vector model" is adopted, with a kernel matrix included for modeling the similarity between XML elements. Our formulation allows individual XML elements to have their own weighted contribution to the overall document similarity while at the same time allows the between-element similarity to be captured. An iterative algorithm is derived to learn the kernel matrix. For performance evaluation, the ACM SIGMOD Record dataset as well as the CEDB dataset have been tested. Our proposed method outperforms significantly the traditional vector space model and the edit-distance based methods. In addition, the kernel matrix obtained as a by-product provides knowledge about the conceptual relationship between the XML elements.