| With the rapid development of computer science and technology,the complex relationship in various fields has produced many complex networks,such as social network,information network,biological network and technology network.Link prediction is an important research content in complex networks.It uses all kinds of existing information in a given network to predict the possibility of link between two nodes that do not exist.It is widely used in the fields of friend recommendation system in social networks,predicting the interaction between proteins,pushing off the evolution mechanism of networks and so on.Most of the classical link prediction algorithms are single machine algorithms based on the information contained in the network topology,but the prediction accuracy of these algorithms has limitations and limits the scope of application.In order to solve the above problems,a hybrid link prediction algorithm integrating the idea of integration is proposed in recent years.Therefore,this paper studies the link prediction problem based on single machine algorithm and hybrid algorithm.The main contents are as follows:(1)Single machine link prediction algorithm.In view of the existing link prediction algorithms based on the information provided by the network topology,it is difficult to effectively measure the network structure information without increasing the computational complexity.Therefore,the concept of local network is introduced,and a node structure similarity link prediction algorithm based on relative entropy and node local structure is proposed.In order to prevent the introduction of too much noise information and the decline of prediction accuracy.Firstly,the local structure of nodes is characterized by defining a second-order local network.Then,the structural similarity between nodes is redefined by using the method of relative entropy.Finally,an algorithm based on relative entropy to measure the similarity index of node structure is proposed.Experimental tests on seven real data sets show that compared with other similarity indicators,the algorithm can obtain better effect and more stable performance under the area under the curve(AUC)measurement standard.It also shows that the algorithm is suitable for networks with small average aggregation coefficient and has better performance on large-scale networks.(2)Hybrid link prediction algorithm based on combination rule method.Aiming at the view that the accuracy of link prediction combination algorithm must be better than that of each single machine algorithm,a hybrid link prediction algorithm based on improved principal component analysis model based on information entropy is proposed.Firstly,seven similarity indexes(CN,LP,LHNII,ACT,Katz,SRW and MFI)are determined by random forest as the best feature set.Then,seven similarity indexes are combined according to the idea of combination,and a feature information fusion model based on improved principal component analysis based on information entropy is proposed.Finally,the weight is given to the characteristic information,and combined with the single machine algorithm,a hybrid link prediction algorithm is proposed.Experimental tests on six real data sets show that the algorithm has better stability and accuracy than other stand-alone algorithms and hybrid link prediction algorithms.(3)Hybrid link prediction algorithm based on fitting learning method.Aiming at the possibility of using ensemble learning technology to improve the accuracy of link prediction,a hybrid link prediction algorithm based on ensemble learning is proposed.Firstly,the stand-alone method is analyzed.The 11 similarity indexes(CN,AA,HDI,HPI,RA,PA,LP,Katz,LHNII,ACT and MFI)are divided into four categories(based on common neighbor,based on path,based on random walk and based on matrix forest index)to extract network features for stacking model.The results show that the prediction effect is poor.Then,11 similarity indexes are assembled as the feature set of complex networks,and the combination strategy of "Pearson r and SVM-RFE" is used to select the best features of applicable networks.Finally,four classifiers are integrated into the bagged and stacked models for hybrid link prediction.Experimental tests on six real data sets show that this algorithm has better stability and accuracy than other link prediction algorithms. |
