Multiple-Campaign Ad-Targeting Deployment: Parallel Response Modeling, Calibration and Scoring Without Personal User Information

TL;DR

This paper introduces a scalable, response modeling approach for multiple ad campaigns that predicts user responses without personal data, using a multi-response generalized linear model to optimize large-scale campaign deployment.

Contribution

It presents a novel scalable method for multi-campaign response prediction without user profiles, addressing training, calibration, and deployment challenges at scale.

Findings

Effective multi-response models for hundreds of campaigns

Improved coverage, precision, and recall in response prediction

Potential for high-throughput decision-making in real-time bidding

Abstract

We present a vertical introduction to campaign optimization; that is, the ability to predict the user response to an ad campaign without any users' profiles on average and for each exposed ad. In practice, we present an approach to build a polytomous model, multi response, composed by several hundred binary models using generalized linear models. The theory has been introduced twenty years ago and it has been applied in different fields since then. Here, we show how we optimize hundreds campaigns and how this large number of campaigns may overcome a few characteristic caveats of single campaign optimization. We discuss the problem and solution of training and calibration at scale. We present statistical performance as {\em coverage}, {\em precision} and {\em recall} used in classification. We present also a discussion about the potential performance as throughput: how many decisions can be done per second streaming the bid auctions also by using dedicated hardware.