AI-Driven Marketing Mix Modeling
- Predictive AI
- H2O Driverless AI
Overview
In this webinar, Akhil Sud and Vijay Raghavan from Allergan/AbbVie's marketing sciences team join H2O.ai's Dr. Michael Proctor to explain why they moved their marketing mix modeling from traditional linear regression to a gradient boosting model with Shapley value attribution. They cover the core limitations of regression-based attribution, including its assumptions about linear response curves, ad-stock decay, and channel interactions, and share validation tests and client results showing where the GBM-plus-Shapley approach outperforms. The discussion also covers practical caveats: no model fully accounts for unexpected market events, and results still need to hold up against business judgment.
Key takeaways
- Why traditional linear regression for marketing mix modeling requires assuming a functional form (linear response, geometric ad-stock decay) and manually testing interaction terms one at a time, which breaks down once dozens of digital channels are added.
- How the team used an H2O AutoML-based GBM approach with Shapley value attribution, rooted in Lloyd Shapley's game theory work, to model non-linear response curves and channel interactions without pre-specifying them.
- How controlled tests with known "correct" answers showed the GBM-plus-Shapley method matched linear regression on purely linear relationships but was substantially more accurate on non-linear and interaction-driven scenarios, for example correctly attributing 540 out of 540 coins of impact versus regression's 503.
- That building a linear marketing mix model took the team two to three weeks per model, versus under a third of that time with the GBM approach.
- How the approach let the team break results down to a granular level, by geography, customer segment, or individual healthcare provider, revealing shifts like speaker programs moving from in-person to virtual impact over time.
- Why both presenters caution that model output still has to be sanity-checked against business experience, citing an example where an unaccounted-for event (a soccer championship sponsorship) inflated an apparent radio ad return.
Start owning your AI