Research Of Tensor Decomposition Methods In Social Tag Recommendation

Social tagging system is an application system which provides a label marking function to the web users. With the Web2.0 is more and more popular, the content-sharing system in which social tagging mechanism is the main function gets rapid development. There are several typical application such as Delicious which shares web pages, Flickr which shares pictures, Last.fm which shares music and songs, and CiteULike which shares academic papers, etc. Social tagging mechanism allows users collaborating with each other to give some labels to the shared resources on the Web site through an open platform. These labels are generally short and full of personality, so that will facilitate the sharing and effective management of resources.Tag recommendation is an important part of the social tagging system, which can automatically provide a user a list of tags that he may be interested in or related for him to choose to use when he labels with tags. Tag recommendation allows users to operate without the trouble of having to input manually, gathers the intelligence of the public users in the network, and provides tags to recommend which can best fit the users'interest or characteristics, so it can greatly facilitate the user's actions and improve the quality of the tags.Using tensor methods in the social tag recommendation algorithms is the latest research in recent years. After analysis of the existing tag recommendation methods with tensor decomposition,however, we found that most of these methods can not effectively deal with the extreme esparseness of the social tag datas or the large number of missing values. In allusion to this defect, we propose a new low order tensor decomposition algorithm, in which we use low order polynomials to decompose the tensor structure of the social tag datas. The low order polynomials which we use mainly includes 0-order,1-order and 2-order polynomials. Experiments show that this method can effectively solve the extremely sparseness of the datas and the problem of missing values, while the performance of precision and recalling has also been improved effectively.Social tag datas are often described as a tripartite hypergraph model, which is intuitive and can express the relationships among each dimension of the tag datas. However, there is always a problem of semantic loss in the dimensions conversion process. In response to this defect, this paper presents a new tripartite tensor decomposition algorithm, in which we use tensor method to decompose the tripartite graph structure. The two-dimensional relationship matrixes produced in the decomposition contains not only the direct relationship between two corresponding dimensions, but also the original relationship information simultaneously exists among the three dimensions, and the datas information they can express is more unabridged and more accurate. So this model can effectively solve the problem of semantic loss.