Keyword Optimization in Search - Based Advertising Markets

Search-based advertising is a multi-billion industry which is part of the growing electronic commerce market. In this work, we study the search-based advertising market from the advertiser’s point of view. There are three natural participants in the search-based advertising: the advertisers, promoting their products to consumers based on their search queries, the users, which are searching for content, and the search providers, who match the advertisers and users by placing ads in search result pages. It is customary today for the advertisers to pay only for ad clicks. This guides both the advertisers’ and the search providers’ strategy. The advertisers’ strategy, which is the focus of this work, requires them to select keywords, bids and a budget constraint imposed on their advertising spending. We abstract an optimization problem in which the advertisers set a daily budget and select a set of keywords on which they bid. We assume that the cost and benefit of keywords is fixed and known, sidestepping this important strategic issue and focusing on the keyword optimization problem. The advertisers’ goal is to optimize the utility subject to its budget constraint. Clearly the advertisers would like to buy the most profitable keywords, subject to the budget constraint. The problem is that there is uncertainty regarding the type and number of queries in a day, and the advertisers have to fix a single policy for the online day. If too few keywords are selected, the advertisers remain with unused budget. If too many keywords are selected, at the time keywords associated with ’good’ ads appear, the daily budget may have already been exhausted. We study the advertisers’ keyword optimization problem in three different settings: in an offline problem setting in which all problem parameters are known beforehand, in a stochastic model in which the advertiser knows only some of the parameters of the stochastic model, and in an adversarial model which makes no statistical assumptions about the generation of the query sequences. For each of these models we provide lower and upper bounds to the performance of learning algorithms while using the notion of regret minimization [1]. We support our theoretical results with extensive simulations in a simulated environment while comparing our algorithm’s performance against the performance of algorithms suggested in [3].