Optimizing on-line display-ad allocation subject to advertiser budget constraints

Abstract

It is common among on-line publishers to monetize their visitors bydisplaying advertisements. To do so, they have the option to use systems called display-ad exchanges to help decide which advertisements are shown to each visitor. The key challenge is to allocate advertisements to viewers in a real time setting. This thesis develops a model that optimizes how the displayad exchange spends the budget of advertisers in order to maximize the revenue of the publisher. This problem is virtually unaddressed in literature. The model is constructed by combining an off-line linear programming model with a linear regression model for web traffic prediction. This combination renders a solution from which it is possible to measure return-on-investment values that can be used by the display-ad exchange to increase the publisher revenue. The thesis develops a greedy Baseline algorithm that simulates key characteristics of a real display-ad exchange. Comparing the return-oninvestment heuristic with the Baseline for a set of real data, shows a 2% increase in publisher revenue. This increase is achieved by spending advertiser budgets more efficiently. The off-line linear programming model shows theoretical revenue improvements in the region of 6%, and that this figure depends on how many advertisers completely consume their budgets.