The Effects Of Bipartite Network Structure On Recommender Systems

The continuous development of the internet technology makes the explosion growth of information in the network.This phenomena leads to users cannot rapidly and accurately finding the information that they're interested in,which knows as information overload.Recommenders are highly effective in solving the problem of choice explosion for online users.The thesis focuses on studying the relationship between the bipartite structure and recommendation performance in recommender systems using theories and approaches of complex network and data mining.The main results are organized as follows:1.In order to explore the relationship between bipartite network's clustering coefficient and recommendation performance,we analyze the performance of various evaluation metrics for six recommender systems in bipartite network that owns different closeting coefficient.The results indicate that the accuracy of the recommender systems are strongly connected with clustering coefficient,the performance of recommender systems in the bipartite networks with high clustering coefficient are more accurate than recommender systems in the low clustering bipartite network.2.This paper focuses on studying the effects of two characteristics that represent the overall bipartite structure features on recommender's performance.Take account of users and items change in recommender system,a new bipartite characteristic(i.e.user-item ratio)is proposed which measures the difference between users and items.The relationship between five metrics with two typical characteristics,i.e.,network density and user-item ratio,are comprehensively analyzed and researched in our work.The results demonstrate that the accuracy evaluation metrics increase while the bipartite network becomes denser,but they do not have a liner relationship.Meanwhile,network density barely affects a recommender's diversity and novelty.In addition,the performances of recommenders become much more accuracy while the user-item ratio becomes larger.The diversity of recommender shows a different status in the bipartite network with various user-item ratios.Furthermore,recommender's novelty is positively linearly related to user-item ratio.3.The influence of community on recommenders has been studied.A new approach is proposed which takes into account the contribution of community structure and introduces the factor of overlap to optimize the accuracy of the similarity between users.The results show that our method is performed better than traditional collaboration filtering algorithm in terms of accuracy,diversity and novelty.