Learning to Rank for Retrieval and Recommendation

Automated systems which can accurately surface relevant content for a given query have become an indispensable tool for navigating large and complex data collections which grow in number every day. At the core of these retrieval systems is the ranking task, i.e., ordering a set of items by their predicted relevance to a query. In recent years, one prominent approach to solving the ranking problem has been learning to rank, in which machine learning methods are employed to learn predictive models that can generate good rankings. This dissertation proposes machine learning algorithms for efficient and effective retrieval of relevant content with a focus on two problem settings: query-by-example retrieval and collaborative filtering with implicit feedback.In the first part of the dissertation, two algorithms that learn a Mahalanobis distance metric optimized with respect to a ranking loss are proposed. The first method suppresses noisy data dimensions effectively during training, while the second method scales to large, high-dimensional data sets while attaining strong retrieval performance on top-of-the-ranking performance metrics. In the second part of the dissertation, a new performance measure for evaluating the recommendation quality of collaborative filtering recommender systems that utilize implicit user feedback is proposed. A feature extraction-based recommender system is then developed and optimized for this performance measure, with extensive empirical evaluations conducted to demonstrate the efficacy of the proposed approach