Introduction

Online advertising is a multibillion dollar industry where competitiveness is directly dependent on accuracy, scalability and sophistication of machine learning algorithms involved. However, machine learning is a crucial but not the only challenge in online advertising. Some of the other problems include

  • Optimization and control: Advertiser typically have a budget and a large set of ads. The system needs to decide which ads need to be run, how much to bid in order to optimize conversions and or profits subject to budget and time constraints. Techniques such as model based throttling have proven useful in addition to feedback based control.

  • Dynamic pricing: While advertisers attempt to determine optimal bids, exchanges have to determine optimal pricing for each ad-impression. Stability considerations have lead to the adoption of generalized second price and VCG auctions for each impression. Pricing impressions differently based on its position, appearance and audience is an open topic of research.

  • Explore/Exploit trade-off: Obtaining training data for building models requires advertiser to spend money. It is important to determine when and how much to spend on exploration vs. trying to exploit known data to maximize profit.

  • Market modeling: When bidding on exchanges, it is important to estimate traffic volume and cost to compute the optimal bid. Such models are useful to determine if spend/profit goal can be reached at all. Economic constraints dictate the choice of models in addition to accuracy.

  • Lifetime value estimation: Advertising involves modeling a funnel of user conversions from views, clicks, conversions, purchases etc. Data from these stream arrives at a different rate with final counts becoming available many days or weeks after the initial event. It becomes important to handle partial responses  while training models. Sometimes, optimizing on lifetime value and immediate value yield very different models.

  • Scalable machine learning: Online advertising can generate billions of events a day requiring the use of scalable algorithms, computing infrastructure. These advances have been partly driven by economic importance of advertising models.

  • Software Engineering: Building online advertising systems require dealing with data ingestion, transformation pipelines, scalable machine learning systems, deployment of such models and managing lifetime of all the models involved in the system. This represents its own unique set of problems not handled by traditional software engineering practices.

    The goal of this workshop is to discuss how machine learning systems operate within the context of an advertising system.