Incrementality Representation Learning: Synergizing Past Experiments for Intervention Personalization

Abstract

This paper introduces Incrementality Representation Learning (IRL), a novel multitask representation learning framework that predicts heterogeneous causal effects of marketing interventions. By leveraging past experiments, IRL efficiently designs and targets personalized interventions, eliminating the need for extensive testing of numerous potential interventions. To ensure generalization to untested interventions and customers, the IRL model extracts low-dimensional representations of intervention features and customer covariates that are predictive of treatment effects and are generalizable across previously tested interventions. Unlike traditional multi-task learning methods that build separate models for each intervention, IRL uses a unified prediction model across past experiments to enhance generalizability.
We empirically validate our framework in the context of promotional campaigns for consumer packaged goods. By synergizing data from 274 previously conducted experiments, our IRL model not only improves the targeting accuracy of tested interventions but also significantly outperforms existing methods in targeting untested interventions and customer segments, overcoming the generalization challenge in high-dimensional decision spaces and the cold-start problem associated with designing new interventions. Furthermore, we develop a decision framework to identify key design features and customer segments for tailoring interventions. Using our model interpretation tool, we demonstrate how companies can customize promotions to enhance profitability across different customer segments.

Keywords: Heterogeneous Treatment Effect, Multi-task Learning, Representation Learning, Personalization, Promotion, Deep Learning, Field Experiments