南京大学学报（自然科学版）2024，Vol.60Issue(5)：753-762,10.DOI:10.13232/j.cnki.jnju.2024.05.006
基于gnn和X-Net融合的ITE估计方法
ITE estimation method based on fusion of g and X-Net
摘要
Abstract
Causal inference helps people make more rational decision-making plans and has wide applications in fields such as e-commerce and precision medicine,and its performance relies critically on the accurate estimation of Individual Treatment Effect(ITE).The selection bias problem and the sample imbalance problem in the observational data affect the accuracy of the individual treatment effect estimation.For the selection bias problem,existing deep learning methods mainly mitigate it by balancing all the covariates,but balancing the processing-independent noise variables in the covariates can lead to inaccurate estimation of individual treatment effect.For the sample imbalance problem,these methods mainly mitigate it by adding sample weights to the loss function.However,this practice does not effectively improve the accuracy of neural network prediction model.In this paper,we propose a method based on deep representation learning,which jointly induces neural networks to obtain balanced shared representations of non-noise variables in the covariates through gnn and IPM(Integral Probability Metric)networks,and then introduces the X-Net to alleviate the sample imbalance problem.The experimental results on semi-synthetic and real datasets respectively show that our algorithm can improve the accuracy of the model individual treatment effect estimation by mitigating the sample selection bias problem and sample imbalance problem.关键词
潜在结果模型/个体因果效应/深度表示学习/选择偏差/样本数量不一致Key words
potential outcome model/individual treatment effect/deep representation learning/selection bias/sample imbalance分类
信息技术与安全科学引用本文复制引用
赵延新,原泽鹏,翟岩慧,牛瑞琪,李德玉..基于gnn和X-Net融合的ITE估计方法[J].南京大学学报（自然科学版）,2024,60(5):753-762,10.基金项目
国家自然科学基金(62072294,61972238) （62072294,61972238）
