Traditional text mining algorithms mostly represent documents based on their textual content. Nevertheless, in the web environment, the content features of web documents are sometimes missing, misleading and unrecognizable due to the lack of well-controlled authoring styles and other reasons. This paper proposed improved k-means clustering algorithm based on user tag. It first used social annotation data to expand the vector space model of k-means. Then, it applied the links involved in social tagging network to enhance the clustering performance. Experimental result shows that the proposed improved k-means clustering algorithm based on user tag is effective.