Authors:
Sonia Slimani
and
Kaiwen Zhang
Affiliation:
École de Technologie Supérieure, Montréal, Canada
Keyword(s):
Real-Time Bidding, Online Advertising, Publish/Subscribe System, Top-k Filtering, Machine Learning.
Abstract:
Real-Time Bidding (RTB) advertising has recently experienced a massive growth in the industry of online marketing. RTB technologies allow an Ad Exchange (AdX) to conduct online auctions in order to sell targeted ad impressions by soliciting bids from potential buyers, called Demand Side Platforms (DSPs). In the OpenRTB specifications, which is a well-known open standard protocol for RTB, the AdX sends bid requests to all DSPs for every auction. This communication protocol is highly inefficient since for each given auction, only a small fraction of DSPs will actually submit a competitive bid to the AdX. The exchange of bid requests to uninterested parties waste valuable computation and communication resources. In this paper, we propose to leverage publish/subscribe to optimize the auction protocol used in RTB. We demonstrate how RTB semantics can be expressed using content-based subscriptions, which allows for selective dissemination of bid requests in order to eliminate no-bid respon
ses. We also formulate the problem of minimizing the number of bid responses per auction, and propose combining top-k scoring with regression analysis with continuous variables as a heuristic solution to further reduce the number of irrelevant responses. We then adapt our solution by considering discrete machine learning models for a faster execution. Finally, we evaluate our proposed solutions against the OpenRTB baseline in terms of end-to-end latency and total paid price over time efficiency.
(More)