Dissimilarity reconstruction in information recommendation

Abstract

A representation of objects in information recommendation named dissimilarity reconstruction (DSR) is proposed in this paper. DSR tries to simulate the gradually transferring mechanism in people's information evaluation process, capture the structure of a data set and retrieve its intrinsic dimensionality. Dissimilarities between objects are first obtained from Vector Space Model (VSM) and then a low-dimensional space is reconstructed by the nonlinear technique Isomap. In the space, Euclidean distance between the associated vectors of two arbitrary objects well represents the dissimilarity between them in sense of evaluation. Experiment on a data set of user activities at bulletin board systems (BBS) has demonstrated the rationality of this representation.