融合数据增广技术与机器学习算法的个人信用评分研究

陆健健 江开忠

摘 要：为了提高个人信用评分模型算法预测精准率，受视觉领域数据增广思路启发，提出融合数据增广技术与机器学习算法的个人信用评分模型。该模型首先对原始个人信用数据进行数据增广处理，然后基于机器学习分类算法训练一个二分类个人信用评分模型，最后基于公开个人信用数据集，分别建立未经过数据增广和经过数据增广处理后的个人信用评分模型。对比准确率、精确率、召回率、F1得分、AUC值和ROC曲线等6个性能评价指标，结果显示，相较于仅基于机器学习算法的个人信用评分模型，融合了数据增广技术与机器学习算法的个人信用评分模型使得分类性能得到了一定提升，分类准确率平均高出5%。

关键词：数据增广技术;机器学习算法;个人信用评分;分类性能评价指标

DOI：10. 11907/rjdk. 192197 开放科学（资源服务）标识码（OSID）：

中图分类号：TP306文献标识码：A 文章编号：1672-7800（2020）008-0040-04

Abstract： Inspired by data augmentation in computer vision， it is feasible to increase the number of training data and make the data set as diverse as possible so as to improve the accuracy of the model of personal credit scoring. After the data is augmented， the performance of the classification task can often be greatly improved. This paper firstly proposes a personal credit scoring model based on data augmentation algorithm. Based on the data augmentation of original personal credit data， a personal credit model is established based on supervised machine learning algorithm. In the empirical part， this paper builds a personal credit scoring model that has not undergone data augmentation and data augmentation processing based on public personal credit data sets. Six performance evaluation indicators， such as accuracy， accuracy， recall， F1 score， AUC value and ROC curve showed that the classification performance was improved more than 5% by the personal credit scoring model based on data augmentation technology.

Key Words： data augmentation; machine learning; credit scoring; classification performance evaluation metrics

0 引言

近年來，随着人工智能和机器学习算法的不断进步与发展，作为人工智能和机器学习领域的一个典型应用，个人信用评分模型取得了长足进步。个人信用评分问题实质上是模式识别领域的一个分类问题，识别个人信用特征并将个人判断划分为不违约和违约两类。具体做法是根据历史数据样本信息，从已知数据中识别违约及不违约者特征，从而总结出分类规则，构建分类算法模型，用于测量借款个人违约概率，为相关决策者或者决策机构提供决策依据[1]。本文将过去个人信用评分研究算法模型主要划分为三大类：传统数学建模方法[2-6];决策树、K近邻判别（KNN）、聚类、神经网络等单一机器学习算法[7-10];集成算法或者多算法融合[11-15]。……