Media Mix Modeling and Optimization with Bayesian Networks
It is a fallacy to believe that Big Data and Artificial Intelligence alone can produce models that are suitable for media mix optimization. Why? Media mix optimization is a fundamentally causal question, and even the most sophisticated predictive models cannot correctly compute the causal effect of marketing actions. In the first part of our webinar, we present Simpson's Paradox to illustrate this challenge.
Media mix models based on Bayesian networks do not automatically overcome this issue. However, Bayesian networks allow us to directly encode causal assumptions from expert knowledge. With that, we can correctly estimate the causal effect of marketing efforts from historical data, and for simple domains that's adequate. For real-world problems, potentially involving dozens or hundreds of marketing-related variables, encoding expert assumptions is no longer realistic, though.
The second part of the webinar focuses on such a scenario, and we use BayesiaLab to machine-learn a high-dimensional Bayesian network model from available historical advertising and sales data. Additionally, we take advantage of the Disjunctive Cause Criterion for identifying the relevant confounders in our model. As a result, we can perform causal inference and compute the direct effects of each marketing driver on the target variable, i.e., daily sales. With this causal model, we proceed to media mix optimization and employ BayesiaLab's built-in genetic algorithm, taking into account cost functions and potential synergies between channels.
In the final part of the webinar, we extend our model to a dynamic Bayesian network with BayesiaLab's Temporalization function. The machine-learned lag structure now captures the dynamic nature of marketing initiatives, and we repeat the optimization process to generate a recommendation for the media mix that maximizes sales within a given marketing budget.