不平衡数据的关键因素筛选方法

Research on Variable Selection Methods of Imbalanced Data

数据不平衡问题的存在,使得模型倾向于将测试样本判别为多数类,导致少数类的分类效果较差。可以从数据和算法两个角度解决数据不平衡带来的问题,本研究主要关注关键因素筛选时不平衡问题的处理,在数据层面使用基于SMOTE抽样的Group Lasso,算法层面使用了调节阈值的Group Lasso,包括分步调节参数和同时调节参数两种方法。最后在307例亚健康患者的问卷数据上使用三种方法建立"肝郁脾虚"诊断模型。从得到的结果来看,基于SMOTE的方法和同时调参的方法得到模型预测效果在灵敏度和特异度上较好。

The existence of data imbalance makes a model tends to predict samples as majority class, resulting in a poor classification effect. The problem of data imbalance can be solved from two aspects of data and algorithm. This research mainly focused on processing imbalance problem in variable selection. In the aspect of data, Group Lasso logistic based on SMOTE sampling was used. In the aspect of algorithm, the Group Lasso with threshold adjusting which include adjusting the parameters step by step and adjusting the parameters simultaneously were used. Finally, the diagnosis model of"liver depression and spleen deficiency"in 307 sub-health patients. questionnaire data was established by three methods. The results showed that the method based on SMOTE and method of simultaneous parameter adjustment have a better prediction in accuracy and sensitivity.