Increasing the robustness of uplift modeling using additional splits and diversified leaf select

Abstract

AbstractWhile the COVID-19 pandemic negatively affects the world economy in general, the crisis accelerates concurrently the rapidly growing subscription business and online purchases. This provokes a steadily increasing demand of reliable measures to prevent customer churn which unchanged is not covered. The research analyses how preventive uplift modeling approaches based on decision trees can be modified. Thereby, it aims to reduce the risk of churn increases in scenarios with systematically occurring local estimation errors. Additionally, it compares several novel spatial distance and churn likelihood respecting selection methods applied on a real-world dataset. In conclusion, it is a procedure with incorporated additional and engineered decision tree splits that dominates the results of an appropriate Monte Carlo simulation. This newly introduced method lowers probability and negative impacts of counterproductive churn prevention campaigns without substantial loss of expected churn likelihood reduction effected by those same campaigns.