Profit maximization through budget allocation in display advertising

Abstract

Online display advertising provides advertisers a unique opportunity to calculate real-time return on investment for advertising campaigns. Based on the target audiences, each advertising campaign is divided into sub campaigns, called ad sets, which all have their individual returns. Consequently, the advertiser faces an optimization problem of how to allocate the advertising budget across ad sets so that the total return on investment is maximized. Performance of each ad set is unknown to the advertiser beforehand. Thus the advertiser risks choosing a suboptimal ad set if allocating budget to the one assumed to be the optimal. On the other hand, the advertiser wastes money when exploring the returns and not allocating budget to the optimal ad set.

This exploration vs. exploitation dilemma is known from so called multi-armed bandit problem. Standard multi-armed bandit problem consists of a gambler and multiple gambling-slot machines i.e. bandits. The gambler needs to balance between exploring which of the bandits has the highest rewards and simultaneously maximising the reward by playing the bandit having the highest return. I formalize the budget allocation problem faced by the online advertiser as a batched bandit problem where the bandits have to be played in batches instead of one by one. Based on the previous literature, I propose several allocation policies to solve the budget allocation problem. In addition, I use an extensive real world dataset from over 200 Facebook advertising campaigns to test the performance impact of different allocation policies.

My empirical results give evidence that the return on investment of online advertising campaigns can be improved by dynamically allocating budget. So called greedy algorithms, allocating more of the budget to the ad set having the best historical average, seem to perform notable well. I show that the performance can further be improved by dynamically decreasing the exploration budget by time. Another well performing policy is Thompson sampling which allocates budget by sampling return estimates from a prior distribution formed based on historical returns. Upper confidence and probability policies, often proposed in the machine learning literature, don’t seem to apply that well to the real world resource allocation problem.

I also contribute to the previous literature by providing evidence that the advertiser should base the budget allocation on observations of the real revenue generating event (e.g. product purchase) instead of using observations of more general events (e.g. clicks of ads). In addition, my research gives evidence that the performance of the allocation policies is dependent on the number of observations the policy has to make the decision based on. This may be an issue in real world applications if the number of available observations is scarce. I believe this issue is not unique to display advertising and consequently propose a future research topic of developing more robust batched bandit algorithms for resource allocation decisions where the rate of return is small.