The era of "big data" has brought marketers and advertisers the opportunity to base decisions on predictive models fueled by massive data on consumer behavior. In this talk I will discuss a series of results from applying machine learning methods to fine-grained consumer behavior data, for real marketing and advertising applications. The consumer behavior data I will focus on includes visiting specific merchants (offline) and specific web pages (online). I will show how predictive performance scales differently with fine-grained behavior data than with traditional data used for targeted marketing; this has important implications for the return firms should expect to get from investing in different sorts of "big" data assets. I will describe a new "sort-of-supervised" method of dimensionality reduction to coalesce related behaviors prior to learning predictive models. And I will illustrate the machine learning concept of "transfer learning" in action, since in this application the ideal data for learning models are quite expensive, but there are useful data that not.
***Foster Provost is Professor and NEC Faculty Fellow at the NYU Stern School of Business, and coauthor of the best-selling O'Reilly book Data Science for Business (http://data-science-for-biz.com). He previously was Editor-in-Chief of the journal Machine Learning, Program Chair of the ACM KDD conference, and his research has been the basis for several successful marketing-oriented companies. Prof. Provost's work has won (among others) IBM Faculty Awards, a President's Award at NYNEX Science and Technology, Best Paper awards at KDD, and the 2009 INFORMS Design Science Award for Social Network-based Marketing Systems.
Catherine Copetas, copetas [atsymbol] cs.cmu.edu