Multi Task Tabular Deep Learning for Propensity Optimization in Marketing

Abstract

RE•WORK Toronto AI Summit – To drive usage growth in any products or services, identifying the proper audience to target for marketing initiatives is a crucial subject. Modern marketers have been leveraging emerging data-driven methods to forecast how well a particular product will fit a certain client. These approaches are typically focused on point estimates of a single product or service, reflecting the fact that marketing efforts are frequently siloed with little consideration given to alternative business lines that are potentially applicable. To move beyond separate predictions, we present a multi-task solution to estimate propensity for three different products and services. This propensity optimization model applies a deep learning architecture for structured datasets to form an ensemble that creates a robust propensity scoring pipeline for millions of clients while producing uncertainty values for reliability. The system can learn single and cross-product knowledge that validates existing business understandings in marketing practices after training on client-level data features. Propensity optimization not only serves as a data-driven proxy to target potential clients in multiple business lines simultaneously, but also paves a novel way to understand and methodically experiment with marketable populations by focusing on propensity-ranked clients within an uncertainty range provided by our model.