Research On Link Prediction Of Weighted Networks

In the study of complex networks,weighted networks are widely used for modeling because they can portray realistic application scenarios more realistically,and link prediction is also an important tool for studying network structure and evolution mechanism because it can predict hidden or possible future network connections.However,there are some problems in the existing research on link prediction of weighted networks that need further improvement:(1)the traditional classical link prediction methods based on network structure only start from the perspective of structural similarity,ignoring the influence of node importance on new connections;(2)the popular link prediction methods based on graph embedding only use natural weight for embedding,ignoring the important role of topological weight in weighted network inscription.To this end,this thesis focuses on the following two aspects of weighted network link prediction research:(1)An unsupervised link prediction algorithm fusing node importance is proposed.In this algorithm,the degree centrality is innovatively selected to calculate the importance scores of node pairs and the node importance information is added to existing algorithms.Custom parameters can be adjusted from both structural similarity and node importance perspectives to achieve optimal prediction performance.In this thesis,5 real undirected weighted network datasets are selected and 11 similar link prediction methods based on network structure are used as benchmarks for comparison experiments.The results show that the algorithm is able to predict weighted networks in real time using existing experience.Compared to similar methods,the indicators such as AUC are improved.(2)A semi-supervised link prediction algorithm fusing topological weight is proposed.In this thesis,a connection strength coupling weight is innovatively designed in the adjacency matrix calculation of weighted networks,which can consider both natural and topological weight in graph embedding methods.At the same time,the algorithm considers the high-order transition probability matrix.Higher-order information is also included in the global representation.In this thesis,8 real undirected weighted network datasets are selected and 9 link prediction methods based on graph embedding are used as benchmarks for comparison experiments.The results show that the algorithm is a general prediction algorithm applicable to different types of network datasets,and the generality is significantly better than that of link prediction methods based on network structure.The indicators such as AUC are better than similar methods due to the effective use of global topological structure information.