Guest Speaker: Deepayan Chakrabarti
Talk Abstract: From humble beginnings a decade or so ago, online advertising has rapidly grown into a huge business, with revenues of 40 billion dollars last year. The driving forces behind its rise have been technological advances in learning from huge datasets, as well as the ability to serve ads and content at massive scales with low marginal costs. In this talk, I will discuss these two aspects in the context of personalization at the per-user level. Specifically, we will look into how the social network of users can be leveraged to quickly learn personalized models for estimating click-thru rates of users on ads and other content. We will discuss models, analyze their properties, understand why some work while others do not, and present scalable inference methods for estimating the (potentially billions of) user-level parameters in such models. Our methods are also broadly applicable to settings with known relationships between features, such as categorical features (e.g., age "buckets") obtained by binning of continuous-valued features.
About the Speaker: Deepayan Chakrabarti graduated from Carnegie Mellon University in 2005, having worked on large-scale graph mining with Prof. Christos Faloutsos. He then worked as a Research Scientist, first at Yahoo! Research (until 2012) and at Facebook since. His interests span a wide range of fields in Big Data, including graph mining, learning and optimization in online advertising, multi-armed bandits, Web search, and Information retrieval. He has about 40 peer-reviewed publications and 20 patents, and is the co-author of a book on graph mining. He has presented tutorials on online advertising at KDD 2009 and CIKM 2008. His work on the R-MAT graph generator is at the heart of the Graph500 supercomputing benchmark, and his COLT 2010 paper received the best student paper award.
Host: Spyridon Blanas
* Deepayan Chakrabarti is a CSE faculty candidate