基于贝叶斯网稀疏结构学习的信用影响因素结构发现

Structural Discovery on Influencing Factors of Personal Credit Based on Sparse Structure Learning of Bayesian Networks

面对海量高维信用数据,传统贝叶斯网络在刻画变量复杂结构和概率关系时遇到了挑战。尝试将基于multi-logit回归的离散贝叶斯网络稀疏方法用于个人信用影响因素结构关系的发现,实现从多维变量复杂关系中抓取重要结构关系;基于解路径探讨了用于结构发现的稀疏贝叶斯网络模型的选择标准,并比较了稀疏贝叶斯网络与经典贝叶斯网络结构学习的性能;结合领域先验知识进一步改进贝叶斯网络结构,定性分析多维变量存在的主要结构关系;在确定多维变量稀疏网络结构的基础上,采用贝叶斯后验估计获取模型参数,并利用贝叶斯网络推理定量分析关键变量对信贷客户类型的直接或间接影响。

Facing massive,high-dimensional data of personal credit,traditional Bayesian networks become much more challenging in representing complex structure and probabilistic relationship among variables.Sparse method of discrete Bayesian networks based on multi-logit regression is applied to dig structural relationships of influencing factors of personal credit and extract important structural relationship contained in multidimensional variables quickly.Then,when the goal is structural discovery,we discuss model selection criteria for sparse Bayesian networks along with the solution path,and compare the performance of structure learning between sparse Bayesian networks and classical Bayesian networks.We improve the structure of Bayesian network combining prior knowledge and analyze primary structural relationship among variables qualitatively.On the basis of determined sparse network structure about multidimensional variables,we employ posterior Bayesian estimation to obtain estimators of parameters,use Bayesian network inference to quantitatively analyze direct or indirect influences of different factors on types of credit customers.