Integrating Customer Segmentation, Predictive Modelling, and Uplift Modelling for Retail Voucher Targeting

Abstract

The rapid growth of e-commerce has encouraged retail companies to optimize data-driven promotional strategies to achieve higher targeting precision and cost efficiency. Mass promotional campaigns without proper segmentation often lead to budget inefficiencies and reduced profit margins. This study aims to optimize voucher targeting strategies for transactions involving growing-up milk products during the January-December 2024 period by integrating customer segmentation, predictive modelling, and uplift modelling. Customer segmentation was performed using K-Means Clustering, resulting in seven distinct clusters representing heterogeneous purchasing behaviors. The clustering output was incorporated as an additional feature in classification models based on XGBoost and CatBoost to predict the probability of targeted voucher redemption. Model evaluation using ROC-AUC, PR-AUC, confusion matrix, and classification report indicated that XGBoost with cluster features achieved the best performance, with a ROC-AUC of 0.9972 and a PR-AUC of 0.5479. Since conventional classification does not measure the causal impact of promotions, uplift modelling with a two-model approach was implemented to estimate the incremental effect of voucher distribution. The results reveal that approximately 1,027 customers (0.46% of 222,036 total customers) belong to the persuadable segment, generating an uplift of 1.91%. These findings demonstrate that uplift-based targeting is more efficient than mass campaigns or probability-based predictive targeting alone. The integration of segmentation, predictive modelling, and uplift modelling provides a more precise and measurable framework for data-driven promotional strategies in retail e-commerce.