WNCG Seminar Series with Deepayan Chakrabarti

Personalized models often revolve around per-user parameters quantifying, say, an individual's interest in a certain product category or susceptibility to a certain type of advertisement, even after known features of the product and the person have been taken into account.  Social networks offer an appealing way to make inferences about such parameters, the intuition being that one's parameter is "close'' to that of one's friends.  We look at this basic scenario from two angles.

First, we consider a Bayesian model that incorporates the social network as a prior, and show that common methods of using the network via Gaussian fields are problematic, especially for real-world social networks with high median degrees. Second, we look at inferring user attributes from partially filled profiles user profiles in a social network, where the standard label propagation algorithm suffers from similar difficulties. In both cases, we find that a model that assumes "partial homophily'' works much better, yielding more accurate inferences and better theoretical guarantees.