基于优化CBR的个人信用评分研究

Optimized CBR for Personal Credit Scoring

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。

Case-based reasoning（ CBR） is proposed to solve sample bias and credit samples dynamic changing in personal credit scoring. CBR fitness issues are analysiszed. Study shows that in Case Retrieval,the weights of variables are the same while the importances of variables in personal credit scoring are different. Besides that,in Case Revise,the weights of similar cases are the same which decreases the effecitves of vote as the useful information is diluted. To solve these problems,a Weight Adjustment Algorithm based on Logistic-BP is introduced in Case Retrieval optimizaion which combins with the high precision of BP and high stability of Logistic. Meanwhile,a Weight Algorithm based on distances is represented in Case Revise to optimize CBR. The empirical analyze reveals that accurancy and explanation are improved and error ratio is reduced. And the model can output the importance of every variable and use the existing personl credit information effectively which is more suitable for personal credit scoring.