Many large datasets associated with modern predictive data mining applications are quite complex and heterogeneous, possibly involving multiple relations, or exhibiting a dyadic nature with associated "side-information". For example, one may be interested in predicting the preferences of a large set of customers for a variety of products, given various properties of both customers and products, as well as past purchase history, a social network on the customers, and a conceptual hierarchy on the products. This talk will introduce a broad framework for effectively tackling such scenarios using a simultaneous problem decomposition and modeling strategy that can exploit the wide variety of information available. The properties and capabilities of this framework will be illustrated on several real life applications.
Thursday, March 24, 2011
Free and open to the public