부산대학교 도서관

In the world of display advertising, advertisers often book a campaign ahead of time and therefore reserve the corresponding supply-inventory. However, they can choose to cancel the campaign any time prior to the start date. When such cancellation happens, it may result in a possible waste of the supply-inventory and a likely increase in opportunity cost. In this paper, we propose a model, referred to as squatting model, which predicts a probability that a given campaign will be cancelled. This probability score along with the supply-inventory requested by the campaign, can then facilitate the decision making process of how many impressions (supply-inventory) should be allocated to this campaign. The term squatting is used because many campaigns reserve inventory without any real intention of using it. Typically, there is no cost for booking and cancelling campaigns, and there is only a cost for utilizing impressions. We used proprietary machine learning framework developed by Microsoft to generate the model and scoring pipeline, deployed it into production environment, and used it to compute the probability-scores on a daily basis for the campaigns that are booked. In this paper, we provide insights into machine learning approach to predict the cancellation of a campaign, experimental results, metrics used, and the applications of the squatting model.