基于多目标进化聚类的信用风险特征识别

Feature Identification for Credit Risk Based on Multi-Objective Evolutionary Clustering

在信用风险识别领域,聚类算法常被用于区分不同风险等级的样本并识别风险特征。然而该领域中通常面临高维数据处理问题,导致传统聚类算法存在不适应此类问题的缺陷:易陷入局部最优、受冗余特征干扰、鲁棒性不强等。采用高维信用风险数据,研究上市公司信用风险,建立信用风险特征识别的三目标优化模型,设计基于分解的多目标子空间聚类算法进行求解。通过算法的横向对比实验,展示了所提出的算法在聚类精度和鲁棒性方面的优势,并根据聚类算法的权重分配结果,归纳总结上市公司信用风险评估过程中应重点关注的指标。

In credit risk identification,clustering algorithm is often used to identify the feature of risk,by distinguishing samples from different inner structure.However,credit risk data has high-dimensional characteristics.When dealing with credit risk data,there often exist flaws in conventional clustering methods,such as trapping in local optimum,being deteriorated by redundant features and poor robustness.Therefore,we look into the credit risk assessment problem,by establishing a tri-objective optimization model and then designing a decomposition-based multi-objective subspace clustering algorithm to solve it.A comprehensive experiment is also conducted to demonstrate the advantages of our proposed method.Then we summarize the key factors for the risk assessment.