Link Prediction Based On Distance Measurement

With the development of large-scale complex networks represented by social networks,the research on complex network analysis has also developed rapidly.Link pre-diction,which is an important research direction in complex network analysis,aims to find the missing links or the possible links in the future using the observed network information.Generally speaking,most researchers usually select a certain kind of struc-tural features in a network to evaluate the similarities between nodes and then predict the connection possibility of node pairs.However,due to the different characteristics of diverse networks,these methods are often unstable in terms of prediction performance.Inspired by k-nearest neighbors(KNN)algorithm,this thesis proposes a new link pre-diction method LP_KNNfor homogeneous networks.LP_KNNcombines several similarity features to improve the prediction accuracy and stability.Further,to solve the prob-lem of link prediction in heterogeneous networks,a link prediction method LP_DISis proposed,which combines the node association strength and meta-path information.(1)LP_KNN:Based on the idea of k-nearest neighbors algorithm,this method fuses several link prediction indexes based on structural similarity to construct feature vectors for node pairs.Then,two sets of k-nearest neighbors of a test node pair were searched in the positive training set and the negative training set,respectively.And then,the cor-responding local mean vectors were calculated.Finally,the distances between the test node pair and the two local mean vectors were combined to get the final prediction score.In this thesis,we find that LP_KNNis not sensitive to the value of k.The experiments show that LP_KNNhas good prediction effect and stability in the experimental networks.(2)LP_DIS:Compared with homogeneous networks,the structure of heterogeneous networks is more complex and contains more network information.LP_DISis proposed for heterogeneous network link prediction.LP_DIScombines the node pair association strength and the similarity measures based on the meta-path and obtains the final pre-diction score by combining the distances of the test node pair to the positive ideal so-lution and the negative ideal solution.In order to verify the prediction performance of the proposed method,several experiments are conducted on multiple DBLP networks.The results show that LP_DIShas good prediction accuracy on heterogeneous networks and fusing multiple features for link prediction is a feasible solution.