构建相对贫困人口识别体系和监测预警机制,增强监测预警机制的针对性和有效性,是相对贫困长效治理的必要条件.由于相对贫困识别的理论和算法研究相对较少,本文提出了一种可解释的相对贫困识别与预警模型,即IEWRP模型.该模型以2018-2020年中国家庭追踪调查问卷(CFPS2018)数据为研究对象,运用方差分析技术对原始数据集进行特征选择.然后通过Gradient Boosting构建IEWRP模型,并与决策树(Classification and Regres-sion Tree,CART)、XGBoost、LightGBM等机器学习算法进行实验对比.最后,结合SHapley Additive exPlanation模型对影响相对贫困识别的相关特征的必要性和重要度进行了可解释性分析,识别出影响相对贫困识别的主要特征.实验结果显示,IEWRP模型的预测准确率、精确率、召回率、F1值、AUC(Area Under Curve)值分别为89.4%、90.6%、95.3%、92.9%、0.95,在准确率、精确率、F1值、AUC值四个方面分别提升了0.15%、0.28%、0.09%、0.23%.
An Explainable Model for Identification and Early Warning of Relative Poverty
The necessary conditions for long-term governance of relative poverty are to build a relative poor population identification system and a monitoring and early warning mechanism,as well as to enhance the pertinence and effectiveness of the monitoring and early warning mechanism.Due to the relatively limited theoretical and algorithmic research on relative poverty identification,in this paper,we propose an explainable model for identification and early warning of relative poverty,i.e.IEWRP model.This model takes the data of China Household Tracking Questionnaire(CFPS2018)from 2018 to 2020 as the research object,and uses technique of variance analysis to select features from the original data set.Then,the IEWRP model is constructed by gradient boosting,and com-pared with machine learning algorithms such as CART(Classification and Regression Tree),XGBoost and LightGBM.Finally,an in-terpretability analysis is conducted on the the necessity and importance of relevant features that affect relative poverty recognition with the shapley additive explanation model,identifying the main features that affect relative poverty recognition.Experimental re-sults show that the prediction accuracy,precision,recall rate,F1 value and AUC(Area Under Curve)value of IEWRP model are 89.4%,90.6%,95.3%,92.9%and 0.95,respectively.The accuracy,precision,F1 value and AUC value are increased by 0.15%,0.28%,0.09%and 0.23%,respectively.
国家科技期刊平台
NSTL
万方数据
