Ads on the Internet are increasingly sold via ad exchanges such as RightMedia, AdECN and Doubleclick Ad Exchange. These exchanges allow real-time bidding, that is, each time the publisher contacts the exchange, the exchange ``calls out'' to solicit bids from ad networks. This aspect of soliciting bids introduces a novel aspect, in contrast to existing literature. This suggests developing a joint optimization framework which optimizes over the allocation and well as solicitation. We model this selective call out as an online recurrent Bayesian decision framework with bandwidth type constraints. We obtain natural algorithms with bounded performance guarantees for several natural optimization criteria. We show that these results hold under different call out constraint models, and different arrival processes. Interestingly, the paper shows that under MHR assumptions, the expected revenue of generalized second price auction with reserve is constant factor of the expected welfare. Also the analysis herein allow us prove adaptivity gap type results for the adwords problem.
A dominant form of advertising on the Internet involves display ads; these are images, videos and other ad forms that are shown on a web page when viewers navigate to it. Each such showing is called an impression. Increasingly, display ads are being sold through exchanges such as Right-Media, AdECN and DoubleClick Ad Exchange. On the arrival of an impression, the exchange solicits bids and runs an auction on that particular impression. This allows real time bidding where ad networks can determine their bids for each impression individually in real time (for an example, see [24]), and more importantly where the creative (advertisement) can be potentially produced on-the-fly to achieve better targeting [22]. This potential targeting comes hand in hand with several challenges. The Exchange and the networks face a mismatch in infrastructure and capacities and objectives. From an infrastructure standpoint, the volume of impressions that come to the exchange is very large comparison to a smaller ad network limited in servers, bandwidths, geographic location preferences. This implies a bound on the number of auctions the network can participate in effectively. A network would prefer to be solicited only on impressions which are of interest to it, and in practice use a descriptive languages to specify features of impressions (say, only impressions from NY). However this is an offline feature and runs counter to the attractiveness of real time bidding. Therefore the exchange has to "call out" to the networks selectively, simultaneously trying to balance the objective of soliciting as many networks as possible and increasing total value, as well as not creating congestion or situations where solicitations are not answered. This leads to a host of interesting questions in developing a joint optimization framework that optimizes over the allocation objective as well as the decisions to solicit the bids. Specifically, a participant would be solicited for only a predetermined fraction of impressions. Moreover, these solicitations, referred to as "call outs" henceforth, need to be performed in a smooth manner and avoid burstiness. The impressions need to be managed in an online manner, which suggests the use of online algorithms. The burstiness properties suggests using a queueing model. And the overall goal is to optimize objectives such as (expected) welfare, revenue. While each of these issues have been considered in isolation, the overall challenge is to develop a joint framework, which in turn raises interesting questions about the interactions between different parts of the framework.
The call out framework is formally modeled in Section 1.1. The online allocation aspect, with a view that the call out constraints acts as budgets, is reminiscent of the online ad allocation framework for search ads, or the Adwords problem [19,5,7], and its stochastic variants [9,27]. However the call out framework is significantly different, which we discuss below.
The Adwords problem is posed in the deterministic setting where the expected revenue is treated as a known deterministic reward of allocating an impression j to an advertiser i. The call out framework has no deterministic analogue; the rationale of the exchange is that the bids are not known. If the bids were known (or internal) then we would only call out the winners (assuming multiple slots) of the auction, which is the path taken by the Adwords problem. The call-out framework is similar to Bayesian mechanism design [21]. This has some fairly broad conceptual implications.
First, is the notion of “adaptivity gap”, where a policy is allowed to react to realization of the random variables. The analysis of adaptivity gap is the central question in the exchange setting. This is also relevant in the context of search ads and the adwords setting where the revenue is achieved on a click which is a random event. The adwords model uses the deterministic expectation but it is reasonable to allow an algorithm to adapt to this event (consider low click through rates and large bids, such that a payout affects the budget substantially). To the best of our knowledge, no analysis of adaptivity gap exists for the adwords problem but such a result will follow from our analysis. In the call out setting, when optimizing for welfare or total value, the assignment occurs after the bids are obtained, which has considerable gap in comparison to the assignment that assigns before obtaining the realizations (reduces to the expectations).
Second, many objective functions such as generalized second price with reserve (henceforth GSP-Reserve), for one or multiple slots, have a very different behavior in the Bayesian and deterministic settings. The gap between assignment after and before the realizations is more stark in this contextconsider running Myerson’s (or similar) mechanism on the expected bids instead of the distributions.
Note that in GSP-Reserve we announce an uniform reserve price, be
This content is AI-processed based on open access ArXiv data.
