Uplift Modeling with Multi-Intervention Variables based on T-Learner

Abstract

The paper addresses the limitations of traditional meta-learner models in the field of causal inference when dealing with multidimensional intervention variables and proposes an innovative modeling framework based on the traditional T-Learner framework. The paper proposes three different multi-intervention variable intervention strategies, namely the Link Strength Model, the Customer Value Model, and the Comprehensive Intervention Model, and selects XGBoost as the basic selector. The paper analyzes different evaluation indicators after model training, including the Qini curve, AUCC, Average Uplift/CATE, Uplift standard deviation, etc. The research results show that the comprehensive intervention model based on multidimensional comprehensive scoring has the most significant effect, which can effectively analyze the causal relationship between feature values and results under the condition of multiple intervention variables, and verifies the superiority of the proposed method in the estimation of intervention effects. This study provides an effective machine learning solution for causal inference in multi-intervention scenarios and has methodological significance for expanding the application boundaries of causal models.