利用网络数据预测企业失信行为  

Predicting the discredited behavior of enterprises via large-scale investment network

作　　者：周涛 李艳丽 李倩 陈端兵 谢文波 吴桐 曾途 ZHOU Tao;LI Yanli;LI Qian;CHEN Duanbing;XIE Wenbo;WU Tong;ZENG Tu(Big Data Research Center,University of Electronic Science and Technology of China,Chengdu 611731,China;Business Big Data Inc.,Chengdu 610041,China;Union Big Data Inc.,Chengdu 610041,China)

机构地区：[1]电子科技大学大数据研究中心,四川成都611731 [2]成都数联铭品科技有限公司,四川成都610041 [3]成都数之联科技有限公司,四川成都610041

出　　处：《大数据》2018年第5期41-49,共9页Big Data Research

基　　金：国家自然科学基金资助项目(No.61433014;No.61673085)~~

摘　　要：传统的企业信用水平分析方法多从企业规模、经营地、行业类别、注册与实缴资本等特征属性出发,缺少基于海量关联数据的深入分析。为了解决这个问题,采集、清洗了大量数据,建立了包含400多万家企业的有向投资网络,其中存在各类失信行为的企业有近26万家。研究结果显示,企业失信行为存在明显的"网络效应",目标企业的股东或者投资企业若存在失信行为,则目标企业发生失信的风险远远大于平均值。基于此,提出了简单的预测企业失信行为的算法,其精确性远远超过了不考虑网络效应的回归方法。Previous enterprise credit level analysis mainly focused on the features including enterprise size, place of operation, industry category, registration and paid-in capital, and lacked in-depth analysis based on massive data. A directed investment network consisted of more than 4 million enterprises was built up, among which nearly 260 000 enterprises have various discredited behaviors. The results show that there is an obvious "network effect" in the discredited behaviors of enterprises. If the target enterprise’s shareholders or its invested enterprises have discredited behaviors, the risk of having discredited behaviors of the target enterprise is far greater than the average. Based on this, a simple generalized linear regression algorithm was proposed to predict the discredited behaviors of enterprises, which is far more accurate than the regression method without considering the network effect.

关 键 词：失信行为预测 网络效应 企业征信 

分 类 号：TN399[电子电信—物理电子学]