针对不平衡数据的泛化预测和特征选择问题,提出了一种引入MCP惩罚函数的AUC回归模型(MCP—AUCR)。该模型采用考虑所有阈值信息的优化目标函数,具有处理不平衡数据的能力,并具有较好的特征选择效果;在讨论该模型定义与原理的基础上,提出相应的循环坐标下降训练算法,并通过数值模拟研究验证其优良性质;针对中国股票市场机械、设备、仪表板块中的上市公司,构建了基于MCP—AUCR的财务预警模型。研究结果显示:该财务预警模型可以选择出可解释的重要财务指标并进行有效预测,显著优于传统模型。
In this study, we propose an AUC (area under the ROC curve) regression with the MCP (the Minimax Concave Penalty) regularization (MCP-AUCR) to deal with the forecasting and feature selection issues for the imbalanced data. The proposed method can solve the imbalanced issues for the optimization of AUC based target and has a good performance on the feature selection. We discuss the idea of the MCP-AUCR and an iterative coordinate descent algorithm. Numerical studies are conducted to show the good property of the proposed method. And an applied study of the financial early warning system for Chinese listed corporations is analyzed as an illustrative example.