Adaptive Modeling for Real Time Analytics: The Case of "Big Data" in Mobile Advertising

Mobile marketing campaigns are now largely deployed through the intermediaries of demand side platforms (DSPs) who provide a performance-intensive real-time bidding (RTB) version of predictive analytics as a service. Performance thresholds are roughly 100ms for DSPs to decide whether and how much to bid for a potential client to receive a particular advertisement via their mobile device. This decision requires simultaneous access to multiple very large databases with typically millions of rows and the ability to execute multiple predictive models (e.g., Logistic regression) to gauge the customer's propensity to engage. In this environment, analytic modeling must be automated via model feedback loops which adjust the models dynamically as real time data streams in. We call this mode of analytics adaptive modeling. We detail the process of adaptive modeling from the perspective of a DSP and describe the corresponding model management environment necessary to plan, execute, and evaluate RTB campaigns.

1. Introduction

Programmatic marketing has become the predominant method of conducting advertising campaigns for mobile media. The traditional marketing relationship between publishers and advertisers has given way to a dynamic intermediary for on-line advertising in the form of ad exchanges which establish real-time markets for connecting advertisers with customers. In this on-line world, customers are not identified until they click on a publisher website which sets in motion a request for bids from current ad campaigns to present their ad to this particular individual. This is typically done through demand side platform (DSP) intermediaries who compete for the right to present the ad by submitting real-time bids to a mediating ad exchange. Ad exchanges therefore operate much the same as stock exchanges, and indeed computational advertising in this context, has much in common with high frequency stock trading because of the associated demanding performance requirements involved. The need to evaluate customer propensity via customer scoring coupled with real-time bid optimization must be done in less than half a second of elapsed time. This requires advanced “big data” technology to manage the very large data sets involved as well as the associated propensity models and bid optimization algorithms.