结合代价敏感与集成算法的个人信用评估模型  

Personal Credit Evaluation Model Combining Cost Sensitivity and Integrated Algorithm

作　　者：张怡 罗康洋 谢晓金 ZHANG Yi;LUO Kang-yang;XIE Xiao-jin(School of Mathematics and Statistics,Shanghai University of Engineering Science,Shanghai 201620,China;School of Data Science and Engineering,East China Normal University,Shanghai 200062,China)

机构地区：[1]上海工程技术大学数理与统计学院,上海201620 [2]华东师范大学数据科学与工程学院,上海200062

出　　处：《软件导刊》2022年第4期186-191,共6页Software Guide

摘　　要：针对个人信用数据存在连续型和离散型交织并存以及类不平衡问题,为提高信用评估分类效果,提出一种结合代价敏感和集成算法的个人信用评估分类模型。通过集成信息价值、互信息、信息增益率和基尼指数特征,选择算法生成最优特征子集。结合代价敏感构建以L1-逻辑回归、弹性网-逻辑回归、贝叶斯、决策树和神经网络为基模型的集成模型,并辅之动态加权投票策略。实证结果表明,将集成特征选择算法的模型指标G-means和F-value与原始特征集相比,分别提升了8%和18%,与mRMR特征选择后模型的预测性能相比也有一定提升,且该模型具有一定的鲁棒性。In order to improve the effect of credit evaluation classification,a personal credit evaluation model combining cost sensitivity and integrated algorithm is presented to solve the problem of continuous and discrete co-existence and class imbalance.The optimal feature subset is generated by the feature selection algorithm of integrated information value,mutual information,information gain rate and gini index,and the integrated model based on L1-Logistic Regression,Elastic Networks-Logistic Regression,Bayes,Decision Tree and Neural Network is constructed by combining cost sensitivity,so as to realize the dynamic weighted voting strategy.The empirical results show that,compared with the original feature set,the G-means and F-value of the model after the feature selection algorithm are increased by 8%and 18%.Com⁃pared with the prediction performance of the model after the mRMR feature selection,the model also has a certain improvement,and showed certain robustness.

关 键 词：信用评估 代价敏感 集成学习 特征选择 机器学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]