Link Prediction Method Based On Network Structure And Attribute Information

Link prediction as a fundamental task in graph data analysis,which has a wide range of application value.Not only it can be used to identify fake links and analyze missing data in the network,but also it can be used to predict possible links in the future,such as predicting protein-protein interactions in the biology,products purchase recommendation in social networks and friends recommendation.In theory,link prediction provides help for the evolutionary mechanism of complex networks,and also promotes the evolution of complex networks.At present,the link prediction task faces the following problems:(1)It is very difficult for researchers to design a model that fits all datasets due to the diversity and complexity of real-world networks.(2)With regard to the prediction of links between the nodes outside the observation network and the nodes within the observation network,the prediction ability of these methods is commonly reduced by the over-fitting of known networks.(3)Many practical applications need to summarize and predict the new nodes with only attribute information,but the new nodes have no structure information.Aiming at the defect of existing research,a link prediction method based on network and attribute information is proposed.A multi-layer perceptron is first applied to perform feature transformation,then a linear layer with random initialization and locking weights is introduced to reduce the over-fitting.In addition,both network structure and attribute feature are used to achieve embedding synchronously,while the two kinds of embedding interact with each other through adversarial training,and the solution space can be limited to avoid over-fitting.For new model,the nodes outside the network can also perform the link prediction task with the nodes in the observable network,as well the nodes with only attribute information.Comparative experiments are conducted between new method and several existing state-of-the-art methods,and the results showed that new method achieved different degrees of improvement on several datasets.