Large-Scale Advertising Portfolio Optimization in Online Marketing

The growth of different auction based online advertising platforms such as Google, Facebook, Bing etc. has provided an opportunity for advertisers to create targeted ad campaigns based on “targeting items” including keywords, cookies, websites, and demographic dimensions. The advertiser aims to determine how much to bid on each targeting item in these ad campaigns in order to maximize the return on investment within a specified advertising budget. The online advertising portfolio optimization problem (OAPOP) allows the advertiser to consolidate campaigns, that share a common objective, across ad platforms and formats into a single portfolio while operating under the constraint of an overall advertising budget. This consolidation of various ad campaigns results in an ad portfolio having tens of millions of targeting items that need to be bid for on a regular basis (often multiple times in a day) by large enterprises. This necessitates the need to develop fast algorithms that can deal with the scale implicit to the problem. We formulate the OAPOP as a Multiple Choice Knapsack Problem (MCKP). The MCKP in the context of the OAPOP could have billions of decision variables. We discuss the structural properties of the MCKP and propose an efficient column generation (CG) algorithm to solve the problem. We perform computational experiments on online advertising instances generated based on data collected from Google Adwords Keyword Planner. The computations demonstrate that our CG algorithm significantly outperforms the state-of-the-art linear time algorithm used to solve the MCKP relaxation for the OAPOP. Another significant practical benefit of the CG algorithm over the linear time algorithm is that it generates the complete revenue-to-cost trade-off function as well as the optimal bids at different spend levels in a single pass through the data.