The Research Of Personalized Recommendation System Based On Bipartite Network

Thanks to the development of Internet and the widely used Web2.0, a variety ofe-commerce websites and social network platforms have provided us an enormousamount of information. However, it is difficult for us to search what we need when weface magnanimity information. Personalized recommendation system, therefore, hasbecome an indispensable part of e-commerce. Personalized recommendation system isan effective tool which is used to solve the overload information problem. It is able topredict the user’s current interests without any user’s intervention for its own strongself-learning ability and real-time capability. The booming personalizedrecommendation technology has an extensive application since the beginning of the21century. Currently, personalized recommendation system is widely used in socialnetwork platform, e-commerce, videos, news, movies, music and other various types ofwebsites.Recently, the study of complex network has achieved considerable progress. Thebipartite network can be taken as a typical network. In this article, we intensivelystudied personalized recommendation algorithm based on the bipartite user-objectnetwork. We investigated the effect of degree of objects in recommendation algorithmand found that properly adjusting initial configuration can enhance both therecommendation accuracy and the diversity. According to the individual object degree,we assign a heterogeneous initial resource for each object. Experimental results showthat, the proposed method outperforms the standard heat conduction method by47.33%,and also outperforms an accurate mass diffusion method by24.04%in recommendationaccuracy. Especially, even compared with an excellent hybrid method of heatconduction and mass diffusion, and the original biased heat conduction method, themanifested method further enhances the recommendation in many respects.Further, in the HHM, the correlation resulting from a specific attribute may berepeatedly counted in the cumulative recommendations from different objects. In thisarticle, we design two improved algorithms that can, to some extent, eliminate theredundant correlations by considering the higher order correlations. One way toeliminate the redundant correlations is from the heat conduction point of view. Theother way to eliminate the redundant correlations is from the mass diffusion point of view. Experimental results show that both algorithms can eliminate the correlations ofredundant and further enhance the recommendation accuracy.