Display advertising is the main type of online advertising, and it comes in the form of banner ads and rich media on publishers' websites. Publishers sell ad impressions, where an impression is one display of an ad in a web page. A common way to sell ad impressions is through real-time bidding (RTB). In 2019, advertisers in the United States spent nearly 60 billion U.S. dollars on programmatic digital display advertising. By 2022, expenditures are expected to increase to nearly 95 billion U.S. dollars. In general, the remaining impressions are sold directly by the publishers. The only way for publishers to control the price of the impressions they sell through RTB is by setting up a reserve price, which has to be beaten by the winning bids. The two main types of RTB auction strategies are 1) first-price auctions, i.e., the winning advertiser pays the highest bid, and 2) second-price auctions, i.e., the winning advertiser pays the maximum of the second highest bid and the reserve price (the minimum price that a publisher can accept for an impression). In both types of auctions, bids lower than the reserve prices will be automatically rejected. Since both strategies are influenced by the reserve price, setting a good reserve price is an important, but challenging task for publishers. A high reserve price may lead to very few winning bids, and thus can decrease the revenue substantially. A low reserve price may devalue the impressions and hurt the revenue because advertisers do not need to bid high to beat the reserve. Reduction of ad revenue may affect the quality of free content and publishers' business sustainability. Therefore, in an ideal situation, the publishers would like to set the reserve price as high as possible, while ensuring that there is a winning bid. This dissertation proposes to use machine learning techniques to determine the optimal reserve prices for individual impressions in real-time, with the goal of maximizing publishers' ad revenue. The proposed techniques are practical because they use data only available to publishers. They are also general because they can be applied to most online publishers. The novelty of the research comes from both the problem, which was not studied before, and the proposed techniques, which are adapted to the online publishing domain. For second-price auctions, a survival-analysis-based model is first proposed to predict failure rates of reserve prices of specific impressions in second-price auctions. It uses factorization machines (FM) to capture feature interaction and header bidding information to improve the prediction performance. The experiments, using data from a large media company, show that the proposed model for failure rate prediction outperforms the comparative systems. The survival-analysis-based model is augmented further with a deep neural network (DNN) to capture the feature interaction. The experiments show that the DNN-based model further improves the performance from the FM-based one. For first-price auctions, a multi-task learning framework is proposed to predict the lower bounds of highest bids with a coverage probability. The model can guarantee the highest bids of at least a certain percentage of impressions are more than the corresponding predicted lower bounds. Setting the final reserve prices to the lower bounds, the model can guarantee a certain percentage of outbid impressions in real-time bidding. The experiments show that the proposed method can significantly outperform the comparison systems.
|
![[NJIT logo]](http://archives.njit.edu/vhlib/images/njit_logo.gif){kind=logo}
![[Van Houten Library logo]](images/vhl-top-image.jpg)
