Three Essays on Data-Driven Personalization and Targeting for Marketing Interventions

Abstract

Companies today face significant challenges in personalizing marketing interventions while balancing long-term objectives, adhering to privacy regulations, and managing complex decision spaces. This dissertation addresses these core issues through three interconnected essays and provides practical methodologies for enhancing personalized marketing interventions in today’s data-driven environment. The first essay (Chapter 1) demonstrates and tackles the challenges of optimizing long-term business performance through targeted interventions. It highlights how cumulative unexplained variations in repeated customer behavior can weaken the direct optimization of long-term out- comes. To address this, the essay introduces a surrogate index methodology using short-term signals, coupled with a novel separate imputation strategy to handle the churn and purchase processes driving customer value. Through simulations and a real-world marketing application, the essay demonstrates that the proposed approach significantly outperforms traditional methods in enhancing long-term targeting effectiveness. The second essay (Chapter 2) examines how Local Differential Privacy (LDP) — a strong privacy technique that introduces noise into individual-level data — affects the accuracy of personalized marketing interventions based on Conditional Average Treatment Effect (CATE) predictions. We show that LDP induces heterogeneous and model-dependent errors that hinder accurate personalization. To address these limitations, we propose an honest post-processing method using an unbiased but noisy proxy combined with iterative boosting and a subgroup cross-learning strategy to ensure honesty and mitigate overfitting. Empirical tests demonstrate that our method significantly improves prediction accuracy and treatment prioritization, enabling organizations to achieve effective personalization despite privacy constraints. The final essay (Chapter 3) introduces Incrementality Representation Learning (IRL), a novel multitask framework for predicting heterogeneous causal effects of marketing interventions. By leveraging past experiments, IRL efficiently designs and targets personalized interventions without extensive testing. It extracts generalizable low-dimensional representations of intervention features and customer covariates. Empirical validation using data from 274 promotional campaigns demon- strates that IRL significantly enhances targeting accuracy and effectively generalizes predictions to both known and untested interventions and customer segments, addressing challenges in high-dimensional decision spaces and cold-start scenarios. Additionally, the essay develops a decision framework and interpretation tool to assist firms in identifying critical design features and tailoring promotions for maximum profitability.