【作者】 李辉；

【导师】 彭怡；

【作者基本信息】 电子科技大学 ， 管理科学与工程， 2017， 硕士

【摘要】 互联网金融的发展和个人信用商业化进程的加快,不仅使得人们可以凭借个人信用享受越来越多便利的服务,也给国家完善信用体系提供了新的思路。对提供这些服务的企业来讲,对用户的个人信用风险进行有效的预测,发现潜在的违约用户,是提高风险管理水平,保证服务质量的基础,因此,对个人信用风险预测模型进行研究,有着重要的价值和意义。传统的预测模型已经不能满足当下风险管理的需求,以分类模型为代表的数据挖掘技术成为了构建个人信用风险预测模型的主流技术,面对不同的分类模型,如何选择出能在自己的数据集上有最好表现的分类模型来进行个人信用风险预测成为了企业所关心的问题。本文基于MCDM方法,主要研究了两个问题,一是如何在单个数据集下对分类模型进行评价和选择,二是如何提高分类模型发现潜在违约用户的能力。针对第一个问题,本文把特征空间对模型表现的影响纳入了考虑,使用多种特征选择方法,并结合MCDM方法,构造了一种多空间多准则的模型综合评价框架。在此框架的基础上,本文以美国著名P2P网贷平台Prosper.com的数据集为例,使用TOPSIS方法,综合六个评价指标对五种单分类器模型在个人信用风险预测问题上的表现进行了评价和比较,为企业在特征空间的构建和分类模型的选择上提供借鉴和参考。实验发现,BPNN模型、LR模型以及SVM模型在Prosper.com的个人信用风险预测中有着较好的综合表现。为了进一步提高对违约用户的预测准确率(TPR),本文以这三个分类模型作为基分类器,用SMOTEBagging模型来进行集成学习。个人信用风险预测是数据不平衡问题,SMOTEBagging模型在这种情况下有比传统Bagging模型更好的TPR表现。本文基于AHP方法对其进行了改进,构建了一种AHP-Based Bagging模型,以在不牺牲SMOTEBagging模型整体表现的情况下提高TPR表现。本文先在27个不平衡数据集上对AHP-Based Bagging模型的有效性进行了验证,发现其能以相当于SMOTEBagging模型一半的集成规模,在0.05的置信度下,得到显著优于SMOTEBagging模型的TPR表现,且在AUC和F1-Measure的表现上没有显著变差。然后将AHP-Based Bagging模型应用在Prosper.com的个人信用风险预测中,也得到了比SMOTEBagging更好的综合表现,并且进一步提高了对违约用户的预测准确率。更多还原

【Abstract】 The development of Internet finance and the acceleration of the commercialization of personal credit not only enable people to enjoy more and more convenient services with personal credit,but also provide a new way of thinking for the country to improve the credit system.For the enterprises that provide these service,effective prediction of the user’s personal credit risk and find the potential default user,is the foundation to improve the level of risk management and ensure the quality of service.Therefore,research on personal credit risk prediction model has important value and significance.The traditional model has been unable to meet the current demand for risk management while the classification model represented by the data mining technology has become a mainstream technology to construct the personal credit risk prediction model.With different classification models,how to select the best classification model for personal credit risk prediction in their own data sets have become the concerns of the enterprises.Based on the MCDM method,this paper mainly studies two problems,one is how to evaluate and select the prediction model under a single dataset,and the other is how to improve the ability of the prediction model to identify the potential default customers.To solve the first problem,and considering the feature space can affect the performance of the model,this paper proposed a multi spatial multi criteria evaluation framework which uses a variety of feature selection methods and combined with the MCDM method.Using the proposed evaluation framework and the data from the famous American P2 P loan platform Prosper.com,5 individual classification models are comprehensive evaluated over 6 criteria by TOPSIS and compared with each other,as a demonstration of feature space construction and classification model selection for companies.The experimental results show that the experiment results show that the BPNN,LR and SVM have a good comprehensive performance in the personal credit risk prediction of Prosper.com.In order to further improve the prediction accuracy for default users(TPR),the three classification models are used as the base classifier and SMOTEBagging model is used for ensemble learning.Personal credit risk prediction has imbalanced dataset problem.In such a situation,the SMOTEBagging model has better TPR performance than the traditional Bagging model.In order to get a higher TPR without sacrifice the overall performance,we.improve the SMOTEBagging model base on AHP method,and construct a model name AHP-Based Bagging.We first check the effectiveness of AHP-Based Bagging model under 27 imbalanced datasets and find that under the confidence level of 0.05,the AHP-Based Bagging model can achieve a significantly higher TPR with just half the ensemble size of the SMOTEBagging model,and the performance in AUC and F1-Measure showed no significantly worse.Then,the AHP-Based Bagging model is applied to the personal credit risk prediction of Prosper.com,and gets a better comprehensive performance than SMOTEBagging.In addition,the prediction accuracy of the default users is also further improved.更多还原