DTI: 2012–13 Funded Proposal
Tensor Completion and Preference Measurement for Distilling Recommendations from Big Data
Recommender systems predict a user’s rating of a given item based on past ratings of the user and other users, plus user and item profiles. Context-aware recommender systems utilize additional state information to improve predictions, since user preferences for items are also a function of context (e.g., time and/or location). Whereas matrix factorization is widely used in recommender systems, tensor factorization can naturally incorporate contextual information.
There has been some preliminary exploration in this area, but there exist numerous opportunities for more sophisticated models as well as scalable methods for tensor-based context-aware recommendation. Practitioners of tensor analysis often pioneer new applications, but typically have limited exposure to tensor decomposition theory and methods.
Beyond new and more accurate context-aware tensor recommendation methods, new directions in user choice-based preference measurement will be pursued, using conjoint analysis as a springboard. Specifically, conjoint subspace analysis will be developed to capitalize on the premise that individual user profiles typically lie in an (unknown) low-dimensional subspace; and preference measurement from ‘window-shopping’ click-stream data will be pursued.
Scalability is a serious concern with existing recommendation systems; but the future of recommendation systems hinges on their ability to distill information from really big, diverse, multi-dimensional datasets. The purpose of this seed grant is to help the co-PIs leverage their strengths, reach out to nurture existing collaborations, and forge new links with the data mining community to compete in forthcoming Big Data solicitations.