Problems in Computational Advertising

Abstract

Computational advertising is a multi-billion-dollar industry, yet it has gotten little attention from academic statisticians. Despite this, the performance of this collection of pricing models, keyword auctions, A/B testing, and recommender systems is largely reliant on statistical technique in almost every element of its design and implementation. Online ad auctions and e-commercial logistics are two of the major components of computational advertising. In a real-time bidding scenario, the objective for the former is to maximize expected utilities. The latter is concerned with the development of statistical modeling for dynamic continuous flows. In turn, this leads to a range of various issues, three of which are discussed in this thesis.Chapter 1 briefly introduces the topics of online advertising and computational advertising. Chapter 2 proposes a new method, the Backwards Indifference Derivation (BID) algorithm, to numerically approximate the pure strategy Nash equilibrium (PSNE) bidding functions in asymmetric first-price auctions. The classic PSNE solution assumes that all parties agree on the type distribution for each participant, and all know that this information is held in common. This common knowledge assumption is strong and often unrealistic. Chapter 3 addresses that gap by providing two alternative solutions, each based upon an adversarial risk analysis (ARA) perspective. Chapter 4 extends the previous methodology for Bayesian dynamic flow models of discrete data to real-valued and positive flows. Finally, Chapter 5 presents some concluding remarks and briefly discusses other problems in computational advertising.