基于神经网络规则抽取的个人信用评估模型研究

Study on individual credit assessment model based on Rules Extraction from Trained Neural Networks

本文针对传统个人信用评估体系中的不足,提出了一种基于神经网络规则抽取的个人信用评估模型。通过对已经训练好的人工神经网络隐层激活值进行聚类分析,减少搜索空间,进而抽取出理解性好、简洁的符号规则。从而产生一组可理解的描述,这组描述能最大限度的模拟已经训练好的原神经网络的推理预测行为。使得评价中的人为因素得到弱化,克服了神经网络在个人信用评估中的"黑箱"性缺陷,增强了模型的稳健性和可理解性。

This paper proposes an individual credit assessment model based on rules extraction from trained neural networks according to the defects of individual credit assessment system. Clustering analysis by artificial neural network that has been trained can reduce the searching scale and extract succinct sign rules ot＇ good understanding. Therefore a group of descriptions comes into being, that can be understood and simulate reasoning predict behaviors o f original neural network well trained in a maximum extent. That weakens the human factors in evaluation, and overcomes neural network＇s defects of ＂black box＂ in individual credit assessment to enhance the robustness and intelligibility of the model.