Examine Two Recommendation and Optimization Problems in Online Advertising

Date of Completion

January 2010


Engineering, Electronics and Electrical|Operations Research|Computer Science




This dissertation examines two recommendation and optimization issues in online advertising. One is to optimize the search engine revenue with the presence of advertiser daily budget constraints. A solution framework that combines a novel linear programming formulation and a method to generate serving plans is designed to achieve the optimal search engine revenue. The ad ranking and pricing auction mechanism suggested by the solution framework is different from the generalized second price auction mechanism. The computational results show a significant improvement of revenue over a greedy algorithm allocating to a query all campaigns that have remaining budget following the second price auction mechanism. Decomposition of the large search engine problem into smaller submarkets and the method of handling tail terms make the method practical both in terms of problem size and the solution speed. ^ The second problem this dissertation examines is to recommend the best performing online publishers' display ad inventory for a request for quota (RFQ) from an advertiser. A system architecture is designed and includes the major components of historical online campaign booking and performance data, inventory availability and performance prediction, a pricing engine, a recommendation engine, and an inventory mix optimization engine. Detailed description of applying natural language processing algorithm Probabilistic Latent Semantic Indexing method (PLSI) to “bag of words” representation of campaigns are provided. A linear programming model is designed to find the best inventory mix to optimize campaign objective while satisfying business constraints. ^