Research On Risk Prediction Methods Of Large-scale Unbalanced Guarantee Networks

Production,operation and consumption are important drivers of national economic development,accounting for more than two thirds of GDP in major economies.As an important part of production,operation and consumption,SME loans play an irreplaceable role in promoting economic growth,promoting innovation,increasing tax revenue,absorbing employment and improving people's livelihood.However,the existing bank loan evaluation system lags behind the development needs of small and medium-sized enterprises,and most of the standards are designed for large enterprises.In order to meet the financial security standards of banks,these small and medium-sized enterprises choose to guarantee each other in order to obtain loans.As more and more enterprises are involved,they form a complex security network.This is a double-edged sword for the national economy.On the one hand,these secured loans help businesses to raise funds rapidly during periods of economic growth development.On the other hand,while complex networks can mitigate the risk of corporate defaults in times of economic downturn,they can also lead to large-scale defaults and spread infections.At present,the research on the credit risk of small and medium-sized enterprises still stays in the evaluation of individual enterprises,lack of risk assessment for the whole guarantee network default warning.Therefore,in this paper,we proposes a risk assessment and prediction method for guarantee network,explores and explains the main impact characteristics of enterprise risk default in guarantee networks,proposes a low-dimensional representation method for complex guarantee loans,designs a credit risk prediction and diffusion model,and finally develops a visual risk analysis system for guarantee networks.In summary,the main contribution of this paper can be described as below:1.Automatic feature extraction method and its financial characteristic exploration: we proposes a method of default risk feature extraction,which is based on deep neural network,to find sparse input and mapping layer weights.The result shows it outperforms other classic algorithms in both precision and recall.Finally,we explores to explain the financial behavior meaning of the risk characteristics extracted by the proposed method.2.Binary high-order and hopping walk network embedding methods: traditional network embedding methods are mainly targeted on social networks.We first introduces the differences in structure and statistical characteristics between guarantee networks and social networks.Then,we design the binary high-order network embedding methods to extract the the binary role feature of vertices.Besides,according on the language model and the bridging characteristics of the guarantee network,we propose hopping walk embedding methods.Afterwards,the experimental results show that the two methods have higher accuracy than other network embedding models in the default classification task of the guarantee network.Finally,we explore the characteristics of the credit network structure learned by these two embedding methods.3.Positive weighted k-nearest neighbor classifier: k-nearest neighbor classifier is one of classic classification methods.In the case of unknown data distribution,it is the preferred method for classification research.However,like most machine learning methods,k-nearest neighbors require an approximate balance distribution in different categories.Conventional imbalanced classification methods can be located into three types: sampling based,cost sensitivity and ensemble learning.They are facing the shortcomings of biased sampling and model complexity.In guaranteed loans,the default ratio of credit accounts for only 6% of the normal ratio.Therefore,in this paper,we design an positive weighted k-nearest neighbor classifier,which is easy to implement and naturally support unbalanced classification.Experimental results show that the proposed method achieves the highest accuracy compared with other imbalanced classification baselines.4.Default risk diffusion model: recent studies have shown that the risk in the guarantee network has the characteristics of diffusion,but there is a lack of effective modeling of this behavior.In this paper,we propose a probability graph method to model the default propagation behavior.We divides a company's default probability into the static probability and the diffusion probability.The hyper parameter d is utilized to adapt to the effects of diffusion orders.Finally,experimental results show that the diffusion model achieves higher accuracy in default classification,especially when the order is under 4.5.Visual analysis system of guarantee networks: although there are many visual analysis systems for complex networks,few of them could adopt to guarantee networks risk analysis.Therefore,in this paper,we design a visual analysis system for guarantee networks,including key features of default visualization,risk pattern discovery,network growth process visualization,risk diffusion path and capital flow analysis.Finally,the effectiveness of the system is demonstrated by the domain experts from commercial banks.