Have you ever wondered how do the ads that you see make it to your screen?For an adtech company like Criteo, every time you browse a web page is an opportunity - an opportunity to put a relevant ad in front of you. However, Criteo isn't the only player in this game, so getting that ad out requires winning an auction: just how much is Criteo willing to pay for a given display?In this talk, we will start with an overview of each of the components that go into the computation of the auction price. The component that we are going to focus on in more detail is the prediction that a certain display will lead to a desired action - be that a click, a visit, a sale, etc. As you probably guessed, such predictions are the output of machine learning models. Historically, the models of choice were logistic regression and gradient boosted trees, but more recently various parts of the pipeline have started being replaced with deep learning architectures.Come and learn about the challenges of training machine learning models on highly imbalanced, noisy datasets made up of hundreds of millions to billions of displays, and keeping them up-to-date as the new feature modalities keep rolling in, all while taking into account the complex business needs of the company.