【作者】 王晓燕；

【导师】 吕进；

【作者基本信息】 长安大学 ， 计算机技术（专业学位）， 2021， 硕士

【摘要】 我国金融体系结构逐渐完善,并且居民的消费方式产生了巨大转变,人们对信用贷款的需求越来越多,信用风险已经成为影响商业银行稳定发展的主要因素。个人信用评估模型的建立,可以帮助商业银行高效的处理大量信贷申请,提高评估的预测精度。目前常用的个人信用评估模型存在可解释性差,准确率不高以及稳定性差等问题。论文将针对上述问题展开研究,构建一个在各方面都表现良好的模型。客观科学的数据集对建立一个良好的模型起非常重要的作用。针对论文中使用的个人信贷数据,进行数据处理。在个人信用评估模型中,特征指标也对模型的表现有重要的影响,但特征选择的方法的科学化规范化水平有待进一步提高。针对数据特征之间的相关性,论文研究了皮尔逊相关系数算法及其不足,引入了距离相关系数,在此基础上结合XGBoost算法对特征的重要性得分排序,提出了DC-XGBoost特征选择算法。随后,论文对DC-XGBoost特征选择算法进行了仿真实验,与传统的特征选择算法进行对比,DC-XGBoost算法表现出了更优的性能。在上述对数据集的预处理和特征选择上,使用深度森林算法建立个人信用评估模型。通过对原始深度森林中的级联森林进行优化,在每层级联结构中引入GBDT算法,既保障了每层生成的概率向量准确性不变的同时,也增加了学习器的多样性,从而提高模型的预测准确率。随后构建了基于改进的深度森林算法的个人信用评估模型。最终与常用的逻辑回归、随机森林以及深度森林作对比分析,实验结果验证了与原始的深度森林相比,改进后的深度森林算法提高了2.6%的模型预测精度。更多还原

【Abstract】 The structure of China’s financial system is gradually improved,and the consumption mode of the people has produced a huge change,people’s demand for credit loans is increasing,credit risk has become the main factor affecting the stable development of commercial banks.The establishment of personal credit evaluation model can help commercial banks efficiently deal with a large number of credit applications and improve the prediction accuracy of evaluation.The current personal credit evaluation models have many deficiencies,such as poor interpretability,low accuracy and poor stability,etc.The thesis will conduct research on the above-mentioned problems and construct a model that performs well in all aspects.The objective and scientific data sets play a very important role in building a good model.Data processing is performed on the personal credit data used in the thesis.In the personal credit evaluation model,feature indicators have an important impact on the performance of the model,but the level of scientific and standardized feature selection methods needs to be further improved.Aiming at the correlation between data features,this paper studies Pearson correlation coefficient algorithm and its shortcomings,introduces the distance correlation coefficient.On this basis,combined with XGBoost algorithm to rank the importance of features,and innovatively proposes the DC-XGBoost feature selection algorithm.Subsequently,the paper conducted a simulation experiment on the DC-XGBoost feature selection algorithm,and compared with the traditional feature selection algorithm,the DC-XGBoost algorithm showed better performance.In the above-mentioned preprocessing and feature selection of the data set,the deep forest algorithm is used to establish a personal credit evaluation model.By optimizing the intermediate forest of the original deep forest,the GBDT algorithm is introduced into the linkage structure of each level,which not only ensures the accuracy of probability vectors generated at each level remains unchanged,but also increases the diversity of the learners,so as to improve the accuracy of the model.Then,a personal credit evaluation model based on the improved deep forest algorithm is constructed.Finally,compared with logistic regression,random forest and deep forest,the experimental results show that the improved deep forest algorithm improves the prediction accuracy of the model by 2.6% compared with the original deep forest algorithm.更多还原