Research Of Link Prediction Algorithm Based On Multiple Features

With the advancement of technology,the scale of networks in the real world is increasing,and the complexity of networks is gradually rising,and our lives are surrounded by more and more complex systems.However,when complex systems are abstracted into complex networks for research,the information collected by the network cannot make a determination of the future change trend of the network.Link prediction,as an important branch of complex network research,can use existing network information to predict the future evolution trend of the network or to screen for incorrect connections.Most existing link prediction algorithms consider only from a single feature,ignoring the multiple features contained in the network,such as the network structure and the properties of the nodes themselves.In view of this idea text proposes two multi-feature link prediction algorithms,which work as follows.1.A link prediction algorithm(LP-TFO)that fuses low-order and textual features is proposed.The algorithm integrates the network structure attributes with the text attributes of the nodes to achieve better prediction results.LP-TFO first obtains the network structure feature matrix,which contains the first-order and second-order features of the network,and calculates its node representation vector;secondly,it constructs the node text feature matrix,and decomposes the text features of the constructed network nodes by SVD matrix to obtain the node text feature vector Finally,LP-TFO fuses the two vectors to obtain a feature vector containing network structure information and node text features,calculates the similarity matrix of the final node feature vectors,and uses the similarity matrix to conduct link prediction experiments in three real citation networks.The experimental results show that the LP-TFO algorithm significantly improves the accuracy of link prediction compared with the traditional algorithm.2.The link prediction algorithm that fuses low-order and high-order features(LPILHA)is proposed.The LP-ILHA firstly obtains the target matrix of the network by the matrix decomposition form of Deep Walk algorithm,and obtains the weight matrix containing first-order and second-order neighbors by calculation,and uses them as the low-order feature weights;secondly,the LP-ILHA predicts the link prediction by Katz algorithm with connected edges of Second,LP-ILHA predicts the connection probability between nodes with connected edges by Katz algorithm and uses this weight as the higher-order feature weight;finally,LP-ILHA combines the lower-order features of the network with the higher-order features and then calculates the connection probability between nodes.The experimental results obtained by the AUC measurement criteria show that the LP-ILHA algorithm has a significant improvement in prediction performance compared with the traditional algorithm,and has a better performance in link prediction of real networks.