When Is More Merrier? A Cloud-Based Architecture to Procure Impressions from Multiple Ad Exchanges
研究移动广告公司应使用一个还是多个广告交易所来降低获取展示机会的总成本，并提出选择性竞价策略，通过故意放弃部分竞价请求来节省计算成本。
This study proposes a method to reduce the cost for acquiring impressions for location-based, mobile advertising firms. Such firms act on behalf of advertisers to execute mobile, in-app, ad campaigns. Ad space is sold on ad exchanges that auction impressions one-at-a-time, on a real-time basis. In this paper, we examine whether ad firms should work with one or multiple ad exchanges to minimize the total procurement cost (equal to the cost incurred to acquire impressions plus the computing cost). By working with more ad exchanges, the ad delivery firm can bid lower on each exchange and potentially save on the total procurement cost. However, ad exchanges typically require ad firms to spend a minimum amount on the exchange. Hence, it is not smart to work with an exchange but acquire very few impressions from this exchange. Working with multiple exchanges also incurs a higher computing cost, corresponding to the computing capacity needed to support the bidding architecture. We solve an optimization problem to determine the optimal number of ad exchanges to use to acquire impressions and the optimal bidding policy on each of these exchanges. We also propose a novel, selective bidding strategy where some bid requests are returned with a zero bid. That is, the ad firm deliberately passes on some opportunities to win impressions. Doing so reduces the computing cost (because returning a zero bid expends minimal computing resources). However, the firm needs to bid higher on other opportunities to meet its demand. We find the optimal selective bidding strategy in addition to the optimal number of ad exchanges and the optimal bidding strategy when non-zero bids are returned. Finally, we demonstrate our solution for a real ad firm (Cidewalk) where the firm is shown to reduce its total cost by 33% by working with multiple ad exchanges (instead of working with a single ad exchange) and the use of selective bidding (instead of returning a nonzero bid for every bid request).