Prediction Pathways: A General Boosted Trees Decomposition and Visualization Methodology

Fooji Inc. is a social media engagement platform that runs social campaigns using its “Just-in-time” delivery network to deliver prizes to marketing campaign participants in real time. To better predict the success of a campaign for a given target audience, Fooji invested in training a machine learning model to predict campaign outcomes. In this paper, we provide a graphical analysis of the inner workings of an accelerated boosted trees model trained to predict the outcomes of just-in-time delivery campaigns. This algorithm generates near-instantaneous speed and high accuracy predictions of marketing campaign outcomes. The analysis performs manual and programmatic analysis of graphical decision paths extracted from the trained model. In the specific case of the example model trained for “Just-in-time Delivery” engagement, we are able to identify the core levers for campaign engagement. As a general tool, we demonstrate that the Prediction Pathways methodology for decomposition of the machine learning algorithm is generally useful for machine learning applications.