Collaborative ranking from ordinal data

Personalized recommendation systems have to predict preferences of a user for items that have not seen by the user. For cardinal (ratings) data, personalized preference prediction has been efficiently solved over the past few years using matrix factorization related techniques. Recent studies have shown that ordinal (comparison) data can outperform cardinal data in learning preferences, but there has not been much study on learning personalized preferences from ordinal data. This thesis presents a matrix factorization inspired, convex relaxation algorithm to collaboratively learn hidden preferences of users through the multinomial logit (MNL) model, a discrete choice model. It also shows that the algorithm is efficient in terms of the number of observations needed