基于移动电信数据个人征信模型研究

发布时间：2018-04-30 19:01

本文选题：征信模型 + 电信数据　；参考：《北京邮电大学》2017年硕士论文

【摘要】：随着社会经济的不断发展,信用评价对个人获得更加便捷的服务和社会运行成本降低都有着积极意义。我国个人征信起步晚,目前信用记录覆盖率较低,但需求旺盛。电信数据包含了个人身份信息、消费记录、欠费违约等和信用评价相关度高的数据,具有质量高、用户多和维度丰富等特点,适合用于个人信用评价。但该方面研究工作较少,因此本文对利用电信数据做信用评价进行了探索分析。本文首先对常用征信模型进行了分析比较。基于统计学的模型有逻辑回归、支持向量机和决策树等,基于智能方法的有神经网络等。本文对上述模型都进行了理论分析和推导,并给出了在实际应用中常见问题比如过拟合等的解决方案。然后对电信数据进行理解分析,参考FICO(Fair IsaacCompany)模型并利用相关分析等方法选取了建模需要的特征数据。随后对数据进行了清洗与预处理包括,填补空缺值、剔除异常值、数据离散化和归一化。预处理之后分析数据发现部分特征间存在多重共线性,不符合逻辑回归模型的假设前提,所以对此采用了主成分分析处理,通过降维解决共线性问题。预处理完毕后,对电信数据进行探索分析,更深入的去认识特征的分布规律和与征信的关系。单个模型去做信用评价有时效果有限,本文将机器学习中集成的思想引入,构建基于随机森林的个人征信模型。结合原始电信数据有冗余信息和类别不平衡特点,在单颗决策树的基础上引入分类别有重复采样并随机选择特征子集,训练多个具有差异的决策树,然后将其组合起来预测输出。模型训练后测试表明其预测精度、准确率、召回率和F1均优于常用征信模型。
[Abstract]:With the development of social economy, credit evaluation has a positive significance for individuals to obtain more convenient services and reduce social operating costs. Our country personal credit investigation starts late, the credit record coverage rate is low at present, but the demand is exuberant. Telecommunication data includes personal identity information, consumption record, default of payment and credit evaluation. It has the characteristics of high quality, multiple users and rich dimensions. It is suitable for personal credit evaluation. However, there is little research work in this field, so this paper explores and analyzes the credit evaluation using telecommunication data. In this paper, the common credit model is analyzed and compared. The models based on statistics include logical regression, support vector machine and decision tree, and neural networks based on intelligent methods. In this paper, the above models are theoretically analyzed and deduced, and the solutions of common problems such as over-fitting in practical application are given. Then the telecom data is understood and analyzed, and the characteristic data of modeling is selected by referring to the FICO(Fair Isaac Company model and using correlation analysis and other methods. Then the data is cleaned and preprocessed, including filling the vacancy value, eliminating the outlier value, discretization and normalization of the data. After preprocessing, it is found that there is multiple collinearity among some features, which does not accord with the premise of logical regression model, so principal component analysis (PCA) is used to solve the problem of collinearity through dimension reduction. After preprocessing, the telecom data is explored and analyzed to understand the distribution law of the characteristics and the relationship between the characteristics and credit information. The effect of a single model to do credit evaluation is sometimes limited. This paper introduces the idea of integration in machine learning to construct a personal credit model based on random forest. Combined with the redundant information and class imbalance of the original telecommunication data, this paper introduces subcategory repeated sampling and random selection of feature subsets on the basis of a single decision tree, and trains multiple decision trees with differences. Then combine it to predict the output. The model training test shows that its prediction accuracy, accuracy, recall rate and F1 are better than those of common credit model.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：F626;TP311.13

【相似文献】