基于优化CBR的个人信用评分研究

发布时间：2018-05-01 22:30

本文选题：信用评分 + 案例检索　；参考：《中国软科学》2014年12期

【摘要】：从样本偏差和信用样本动态变化问题出发,以CBR(案例推理)方法建立个人信用评分模型。研究发现,基于CBR的个人信用评分模型在案例检索环节假设特征集中各特征变量具有相同权重,与个人信用评分实际不符;在案例修正环节假设相似案例权重相等,导致已有数据信息无法得到充分利用。针对这些局限性,本文设计了基于Logistic回归-BP神经网络的权重调整算法,结合BP神经网络的高精度及Logistic回归的稳定性计算个人信用评分各特征变量的权重,对案例检索进行优化;设计基于距离的投票算法计算各相似案例的权重,对案例修正进行优化。实证实验证明基于优化CBR的个人信用评分模型精确性和解释性均有所提高,错分率降低,能够输出各指标的重要性,有效的利用已有数据信息,更加适用于个人信用评分。
[Abstract]:Based on the problem of sample deviation and dynamic change of credit sample, CBR (case based reasoning) method is used to establish personal credit scoring model. It is found that the CBR based personal credit scoring model assumes that each feature variable in the feature set has the same weight in the case retrieval link, which does not conform to the actual situation of the individual credit score; in the case revision link, the similar case weight is assumed to be equal. As a result, the existing data information can not be fully utilized. Aiming at these limitations, this paper designs a weight adjustment algorithm based on Logistic regression BP neural network, combines the high accuracy of BP neural network and the stability of Logistic regression to calculate the weight of individual credit score characteristic variables, and optimizes the case retrieval. A distance-based voting algorithm is designed to calculate the weights of similar cases and optimize the case correction. The empirical results show that the accuracy and explanation of the personal credit scoring model based on optimized CBR are improved, the error rate is reduced, the importance of each index can be outputted, and the existing data information is used effectively, which is more suitable for personal credit rating.
【作者单位】：哈尔滨工业大学管理学院;
【基金】：国家自然基金项目(70871030) 黑龙江省自然科学基金项目(G200914)
【分类号】：F832.4

【参考文献】