Research On Network Node Centrality Prediction Based On Link Prediction

As a method to mine and predict missing information in complex networks,link prediction has been paid more and more attention.Although researchers have proposed a variety of link prediction algorithms,there is a lack of systematic research on the difference of prediction accuracy between different link edges.At the same time,in the research of future degree increment of complex network nodes,the current popularity prediction algorithm based on preference attachment can not well solve the problem of popularity prediction of new nodes in the network.This paper aims at clarifying the accuracy distribution characteristics of complex network link prediction and realizing the prediction of node future degree centrality by aggregating the prediction results of links.The main work of this paper is as follows:(1)The traditional link prediction values focus on the average value of the predicted AUC.In this study,the AUC of each edge in the test set was systematically analyzed,and it was found that the predicted results showed significant polarization phenomenon.Through further analysis,it is found that the prediction of the edges between nodes with low degree is generally poor,even lower than the accuracy of random prediction.However,the predicted result of the link between nodes with big degree is very high,close to 1.The research results reveal the heterogeneity of the prediction results of current link prediction algorithms,and provide a direction for improvement of link prediction algorithms.(2)Based on the possibility of the future occurrence of the link edge,an algorithm of prediction of node degree increment based on the link prediction is proposed in this paper.The prediction of node degree increment is realized by aggregating the possibility of the link edge occurrence.Fourteen link prediction algorithms are systematically tested in this paper.Experiments show that the method proposed in this paper can predict node degree increment accurately.In this study,the prediction accuracy of nodes with high degree of discovery is higher,while the prediction accuracy of nodes with low degree is lower.The results also reveal that the prediction accuracy of the high-degree nodes is more reliable than that of the low-degree nodes in link prediction,which verifies the heterogeneity of the link prediction accuracy.(3)Since the recommender system is essentially the link prediction of bipartite graph network,we extend the prediction method of network node centrality to recommender system in this paper.At present,the main method for predicting the popularity of items is preference attachment.Preference attachment is based on the historical popularity of an item to predict the future popularity.Preference attachment usually has good prediction results for nodes with high degree,but it will fail at all for nodes with low degree,especially those nodes with the same degree.In view of these defects of preference attachment,this paper proposes a popularity prediction method based on recommendation based on collaborative filtering algorithm,which predicts the future popularity of items by combining each user’s preference and user’s activity.Experiments show that this method is better than the preference attachment in predicting the future popularity of items with low degree or the same degree.