随机森林在P2P网贷借款信用风险评估中的应用

发布时间：2018-02-26 23:23

本文关键词： P2P网络借贷信用评估随机森林特征选择　出处：《山东大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着互联网金融逐步发展,P2P网络借贷作为一种创新的互联网金融模式在当下变得炙手可热。近年来,凭借着高收益、便捷等理财优势,各类P2P网络借贷平台得到了爆发式增长。然而,我国P2P网络借贷因其较短的发展时间、落后的风险分析技术、以及尚不健全的法律法规体系等使得投资者面临着严重的资金安全隐患问题。信用风险已成为P2P网络借贷行业的首要发展瓶颈,如何为P2P网络借贷者建立一套良好的风险评估模型,显得非常迫切与必要。但是学术界对此还缺少深入的研究,风险评估依然处于传统个人借贷风险评估方法的简单模仿阶段。本文我们选择一种组合分类器模型——随机森林(Random Forest),该算法能够较好的容忍噪声,不容易产生过度拟合,且有着较高的稳定性,相比于传统的单分类器模型,能够更好的处理信贷风险评估问题。文中对随机森林算法的相关理论进行了详细介绍,并在此基础上通过引入代价敏感学习法提出了加权随机森林算法(WRF),进而提高错判代价较高的类别的准确率,增强模型的实用性。在实证阶段,我们首先对数据进行了预处理,主要包括剔除离群值点,补齐缺失值,归一化以及相关性检验;接着,采用五重交叉数据利用RF算法进行特征指标的选择,为实证部分打下了坚实的基础;然后,利用Lending Club开放数据集,以及German、Austria公开银行信贷数据集,建立了以随机森林为基础的个人信用评估模型,并与Logistic回归,KNN,SVM,ANN等传统信用风险评估方法进行对比。通过实验我们知道:不论是P2P信贷数据还是传统银行信贷数据RF模型总体分类效果最佳,这说明RF算法更加适用于构建信用评估模型。最后,我们采用SMOTE算法针对P2P不平衡数据进行了处理,使得分类结果更加具有现实参考意义。
[Abstract]:With the gradual development of Internet finance, P2P network lending, as an innovative Internet finance model, has become a hot issue. In recent years, with the advantages of high profit, convenience and other financial advantages, All kinds of P2P network lending platforms have been exploding. However, due to their short development time and backward risk analysis technology, P2P network lending in China, As well as the imperfect system of laws and regulations, investors are faced with a serious problem of capital security risks. Credit risk has become the primary bottleneck in the development of P2P network lending industry. How to establish a good risk assessment model for P2P network borrowers is very urgent and necessary. The risk assessment is still in the simple imitation stage of the traditional personal loan risk assessment method. In this paper, we choose a combined classifier model, Random Forest Random Forestwood, which can tolerate noise well and is not easy to overfit. Compared with the traditional single classifier model, it can deal with the credit risk assessment problem better. In this paper, the related theories of stochastic forest algorithm are introduced in detail. On the basis of this, a weighted stochastic forest algorithm is proposed by introducing the cost sensitive learning method, which can improve the accuracy of the category with high cost and enhance the practicability of the model. In the empirical stage, we preprocess the data firstly. It mainly includes eliminating outliers, compensating missing values, normalizing and correlation testing. Then, we use five cross data to select feature indicators using RF algorithm, which lays a solid foundation for the empirical part. Using the Lending Club open data set and the open bank credit data set of Germani Austria, a personal credit assessment model based on random forest is established. And compared with the traditional credit risk assessment methods such as Logistic regression KNN / SVMANN and so on. Through the experiment, we know that both P2P credit data and traditional bank credit data RF model overall classification effect is the best. This shows that RF algorithm is more suitable for building credit evaluation model. Finally, we use SMOTE algorithm to deal with P2P unbalanced data, which makes the classification results more practical reference significance.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：F832.4;F724.6

【相似文献】