Learning to Rank: Online Learning, Statistical Theory and Applications.

Abstract

Learning to rank is a supervised machine learning problem, where the output space is the special structured space of emph{permutations}. Learning to rank has diverse application areas, spanning information retrieval, recommendation systems, computational biology and others. In this dissertation, we make contributions to some of the exciting directions of research in learning to rank. In the first part, we extend the classic, online perceptron algorithm for classification to learning to rank, giving a loss bound which is reminiscent of Novikoff's famous convergence theorem for classification. In the second part, we give strategies for learning ranking functions in an online setting, with a novel, feedback model, where feedback is restricted to labels of top ranked items. The second part of our work is divided into two sub-parts; one without side information and one with side information. In the third part, we provide novel generalization error bounds for algorithms applied to various Lipschitz and/or smooth ranking surrogates. In the last part, we apply ranking losses to learn policies for personalized advertisement recommendations, partially overcoming the problem of click sparsity. We conduct experiments on various simulated and commercial datasets, comparing our strategies with baseline strategies for online learning to rank and personalized advertisement recommendation.