The last ten years have seen a tremendous growth in Internet-based online services such as search, advertising, gaming and social networking. Today, it is important to analyze large collections of user interaction data as a first step in building predictive models for these services as well as learn these models in real-time. One of the biggest challenges in this setting is scale: not only does the sheer scale of data necessitate parallel processing but it also necessitates distributed models; with hundreds of million active users on major online services such as Facebook, Twitter, Amazon or Google, any user-specific sets of features in a linear or non-linear model yields models of a size bigger than can be stored in a single system.

In this tutorial, I will give an introduction to distributed message passing, a theoretical framework that can deal both with the distributed inference and storage of models. After an overview of message passing, I will discuss and present recent advances in approximate message passing which allows to control the model size as the amount of training data grows. We will also review how distributed (approximate) message passing can be mapped to generalized distributed computing and how modeling constraints map on the system design. In the second part of the talk, I will give an overview of the application of these techniques to real-world learning systems, namely:

Gamer ranking and matchmaking in TrueSkill™ and Halo 3
AdPredictor click-through rate learning and prediction in sponsored search
User-action models in Facebook's information distribution and advertising pipeline