Research On Recommendation Algorithm Based On Bipartite Graph

Recently, personalized recommender systems have become indispensable in a wide variety of commercial applications due to the vast amount of overloaded information. From 1990 s, lots of researches on recommender system have appeared. This paper makes some improvements for recommendation algorithm based onbipartite graph.Network-based recommendation algorithms for user–object link predictions have achieved significant developments. But most previous researches on network-based algorithm tend to ignore users’ explicit ratings for objects or only select users’ higher ratings which lead to loss of information and even sparser data. With this understanding, we propose an improved network-based recommendation algorithm. In the process of reallocation of user’s recommendation power, this paper originally transfers users’ explicit scores to users’ interest similarity and user’s representativeness. Finally, we validate the proposed approach by performing large-scale random sub-sampling experiments on a widely used data set(Movielens) and compare our method with other two algorithms by two accuracy criteria. Results show that our approach significantly outperforms the original network-based recommendation algorithm.A brief review of the past researches on CF shows that methods for calculating users’ similarities are almost Pearson Correlation or(adjusted) Cosine Similarity. This leads to same recommendations for different users because popular objects or users often win a heavier weight in the process of recommendation. Moreover, it has been increasingly recognized that the gains of the recommendation accuracy are often accompanied by the losses of the diversity. In order to walk out of the accuracy-diversity dilemma, we propose a new method named collaborative filtering based on bipartite graph and random walk with choice which replaces the traditional Pearson Correlation or(adjusted) Cosine Similarity for calculating users’ similarities.This paper transfers different user-object data into weighted graph or unweight graph according to complex network theory. Results show that our approach significantly outperforms the ordinary user-based collaborative filtering method and power-law-based recommendation algorithm in diversity without lowing recommendation accuracy.