Personalized Recommendation Based on Unbalanced Symmetrical Mass Diffusion

Information filtering model for personalized recommendation has dramatically promote the development of recommendation technology. In various kinds of information filtering models, mass diffusion model has rise fruitful researches. Although traditional work assumes in heterogeneous bipartite network consideration of initial resource from the collected object can much enhance the recommendation accuracy, its intrinsic unsymmetrical assumption that similarity is only determined by power of mass resource diffusion from the collected to the uncollected proves defective, adversely limiting further improvement of recommendation accuracy. After investigation, we believe symmetrical diffusions involve not only the direction from the collected object to the uncollected but also the reverse. Moreover, we find the bidirectional diffusions are unbalanced, and comprehensively propose a unbalanced symmetrical diffusion model (USD) to conquer the above drawbacks. Extensive numerical experiments, performed on two benchmark data sets Movielens and Netflix, shows more outstanding improvement than mainstream diffusion based methods, with respect to accuracy, diversity and personalization.