几类群组变量选择方法及其块坐标下降算法  

Several Types of Group Variable Selection Methods and Block Coordinate Descent Algorithm

作　　者：李春红[1] 钟小敏 宗瑞雪 

机构地区：[1]广西大学数学与信息科学学院,广西南宁

出　　处：《应用数学进展》2019年第8期1478-1486,共9页Advances in Applied Mathematics

基　　金：国家自然科学基金资助项目(No.71462002).

摘　　要：在复杂数据中变量往往成组出现,考虑了Lasso、SCAD、Bridge及MCP四种不同模型选择的惩罚项,研究了它们在群组变量中的方法及其块坐标下降算法,在Logistic模型的条件下进行模拟,结果表明Composite MCP组惩罚方法在预测能力和变量选择上均优于其他三种群组惩罚方法,并运用到销售网络办公软件公司的广告数据中,结果表示四种方法中Composite MCP方法在广告转化研究中效果是最优的,通过比较,选择出影响广告转化的群组结构及单个变量,为选择投放策略提供合理的依据。In complex data,variables often appear in groups.Considering the penalty terms of four different models selected by Lasso,SCAD,Bridge and MCP,their methods in group variables and their block coordinate descent algorithm are studied.The simulation is carried out under the condition of Logistic model.The results show that the Composite MCP penalty method is superior to the other three group punishment methods in predicting ability and variable selection,and is applied to the company advertising data for saling network office software.In the method,the Composite MCP method is the best in the advertisement conversion research.By comparing and selecting the group structure and individual variables that affect the advertisement conversion,it provides a reasonable basis for selecting the delivery strategy.

关 键 词：群组变量选择 块坐标下降算法 LOGISTIC模型 广告转化 

分 类 号：F2[经济管理—国民经济]