GCN-based Dynamic Network Directed Edge Weight Prediction

The link prediction of complex networks has always been an important research direction in the field of complex networks.Link prediction includes not only the prediction of unknown edges,but also the prediction of future edges.Extending a complex network from a static network to a dynamic network can effectively distinguish between unknown and future edge predictions,and link weight prediction also extends the presence or absence of link predictions to the positive and negative links and possible links.For transaction systems,especially those with strong anonymity and high risk of fraud,such as those that use Bitcoin for transactions,it is particularly important to have a rough estimate of the reliability of the counterparty in advance.After the evaluation and scoring mechanism for historical transactions has been introduced,it is a good choice to judge the reliability of the counterparty based on the counterparty’s score.However,the transaction score reflects only the transaction history score,and for low-frequency users who rarely conduct transactions and fraudulent users who falsify transaction records and get high scores,it is impossible to use simple historical scores to assist users in determining transaction reliability.Therefore,our goal is to integrate the various attributes of the user and the relationship with other users,and provide users with a possible predictive score of the transaction to assist the user in making judgments.After clarifying our basic tasks,the first thing we thought of was to build our task on a graph model.The users in the trading system are the nodes,and the scoring is the weight of the edges between the nodes,because the scoring is directed and positive and negative.,So the edges are positive and negative,and because we need to predict the future score based on the historical score,we build on the dynamic graph model.Ultimately,our problem is how to predict the weights of edges in a dynamic directed network.The timely rating forecast of customers is crucial for trading service evaluation.Due to the high complexity and variability of rating,traditional static graph models have been proved to be ineffective in learning the topological and temporal features of dependent nodes and the representation of edge weights.In this paper,we propose a real-time trading evaluation system to do the trading transaction prediction on the bitcoin marketplace and a novel deep learning framework to tackle the dependent problem of topological and temporal information.We elaborate a hierarchical structure including dynamic rating construction network layer,extracting node features layer and edge learner layer.The major contributions of this work can be summarized as follows:1.We propose a new deep learning framework to predict the trust evaluation scores of online trading platforms.Aiming at the link prediction problem of dynamic networks,the framework is divided into three layers.The first layer of dynamic network construction is used to use the original chaotic and disorderly The data is organized into an orderly dynamic network;the second layer of node learning layer is used to learn the representation of the nodes of the network;the third layer is the link prediction layer to learn the link prediction of the dynamic network.2.About aiming at the link weight prediction problem of dynamic networks,we added an asymmetric cross-mapping model to learn the link weight prediction of dynamic networks on the basis of the original evolution graph convolutional neural network.3.We finally obtain better results from predictions of edges’ weights than the one from the state of art.At the same time,we achieve the improvement of transaction service evaluation.