Action Attention GRU: A Data-Driven Approach for Enhancing Purchase Predictions in Digital Marketing
Girim Ban, Simon S. Woo, David Sung · 2024
Modeling user behaviors is critical in digital marketing. Digital marketing agencies often operate with logs defined by advertising IDs (ADID). However, ADID data from digital marketing firms often lack rich user profiles and detailed item information. Considering such attributes, we present a data-driven model, the Action Attention bidirectional Gated Recurrent Unit (AAGRU) to effectively learn sequences of user behaviors without explicit knowledge of the actors or targets for conversion prediction. Tailored to predict impending purchases based on ADID's digital journey, AAGRU leverages two pivotal components: the action block and the interval block. The former adeptly captures salient actions in the journey through attention mechanisms, while the latter discerns temporal nuances, such as impulse and deliberate buying tendencies. This tailored approach enables digital marketing agencies to identify latent customers ready for purchase, thus optimizing targeted advertising and conversion strategies. Our experimental results affirm AAGRU's superiority over extant deep learning models. Notably, in simulations, AAGRU shows impressive performance against our company's best audience group. Over two days, the model achieves elevated conversion rates (CVRs) of 14.24% and 15.79%, with stable click-through rates (CTRs) of 1.50% and 1.68% respectively. These CVRs indicate significant enhancements of approximately 137% and 176% over the comparison group's metrics, underscoring the model's potential in refining marketing efficacy.