Personalized prediction predicts a user's preference for a large number of items through user-specific as well as content-specific information, based on a very small amount of observed preference scores. The problem of this kind involves unknown parameters of high-dimensionality in the presence of a high percentage of missing observations. In this situation, the predictive accuracy depends on how to pool the information from similar users and items. Two major approaches are collaborative filtering and content-based filtering. Whereas the former utilizes the information on users that think alike for a specific item, the latter acts on characteristics of the items that a user prefers, on which two kinds of recommender systems Grooveshark and Pandora are built.
In this talk, I will present our recent research on regularized latent-factor modeling and compare with state-of-art recommenders in terms of predictive performance. Special attention will be devoted to the impact of nonignorable missing and social networks on personalized prediction.