| With the rapid development of Internet finance and big data technology,traditional banks are facing huge challenges in their business models.In response to this situation,major commercial banks in China have launched online loan products to attract customers,but these online loan products have resulted in a large number of nonperforming assets due to the lack of effective risk control methods.The traditional machine learning model mainly analyzes the dimensional features of user data to find the difference between normal values and outliers.However,in actual scenarios,user data has hidden behaviors and characteristics,such as emergency contacts,time series.At the same time,as time goes by,graph data will also form new nodes and edges,which makes a large-scale financial social network graph formed between user data.The research content of this paper is as follows:(1)In credit data,each node or edge of sample data has time characteristics,and new nodes and edges will appear over time.This dynamic behavior often conveys important information.This paper proposes a time-series-based the feature embedding method extracts the time series features and adds them to the node matrix,and at the same time embeds the time features,and adds the embedded results into the graph as the attributes of the nodes.Expressive power and performance in graph data.The comparative experiments prove that the generalization ability and performance of the model are improved after the method is processed.(2)In the traditional forecasting model,the characteristics of sample data are mainly analyzed and modeled,which often ignores the correlation between samples.This paper proposes a credit default model of a graph attention network based on feature aggregation.Usually,each user node is affected by neighbor nodes to varying degrees.By introducing an attention mechanism,the coefficients of different neighbor nodes are calculated to strengthen important the influence of neighbor nodes on their own nodes.At the same time,through feature aggregation,the features of adjacent nodes can be effectively collected,and the information and features of adjacent nodes can be continuously fused together to form a new node representation.In this way,as time changes,new nodes can be directly strengthen the representation through aggregation without re-reading the entire network structure,enhancing the generalization ability of the model.Through experimental comparison,the accuracy of the model has been improved to varying degrees.(3)According to the above research methods,a credit risk control system based on deep learning is established,which includes four modules: data import module,data processing module,model training module,and risk visualization module.Through this system,business personnel can import data and independently train through the model,and finally display the training results of various indicators of the model in a visual way,improving the decision-making ability of business personnel. |
