Primary Submission Category: Heterogeneous Treatment Effects
Synergizing Experiments: Designing Personalized Marketing Interventions through Incrementality Representation Learning
Authors: Ta-Wei Huang, Eva Ascarza, Ayelet Israeli,
Presenting Author: Ta-Wei Huang*
In the pursuit of personalization, firms aim to tailor interventions to match individual customer preferences. Traditional methods typically consist of two stages: first, a set of predefined interventions are tested and evaluated across various customer segments; then, the most effective intervention for each segment is assigned. Although this approach provides some benefits of personalization, they often fall short in achieving complete customization, constrained by the capacity to test only a few interventions within a set number of segments. Furthermore, these methods do not provide insights for designing new interventions or for targeting different segments, thereby restricting the scope and potential of data-driven personalization.
This research introduces a novel approach enabling companies to design personalized interventions more efficiently by capitalizing on historical experimental data. We develop a flexible causal machine learning framework, termed incrementality representation learning, which estimates the conditional average treatment effect (CATE) based on intervention characteristics and customer covariates and extracts low-dimensional representations of these features to capture the heterogeneity in treatment effects. This approach allows firms to leverage past experiments to identify the most effective type of intervention for any specific