Quality-Aware Ranking for Recommendation Systems

Abstract: Recommender systems commonly present users with a ranked list of items, known as a top-K answer. These rankings are typically generated by scoring candidate items based on information collected from users. Such scores pose a challenge for ranking items under inherent score uncertainty, which may arise from data unreliability and missing data. In this thesis, we address top-K queries over uncertain data in recommender systems. We explicitly model score uncertainty by representing scores as probability distributions rather than deterministic values. We first study static top-K recommendation under uncertain scores, where the recommendation size is fixed in advance and the answer is generated in a single step. We show that, under score uncertainty, the choice of ranking semantics should be aligned with the target quality measure. Based on this observation, we provide a formal analysis connecting quality measures and ranking approaches, and introduce rank-based methods for generating high-quality recommendations under uncertain scores. We then shift from the static top-K paradigm to an adaptive sequential recommendation setting under uncertain scores, where the answer is constructed incrementally rather than generated as a fixed list in a single step. In this setting, feedback observed after each recommendation can be used to update the recommendation process before selecting subsequent items. We propose adaptive ranking semantics, feedback-based update mechanisms, and stopping criteria that support the generation of high-quality recommendations while allowing the answer size to be determined dynamically. Finally, we study ordinal-aware learning of score distributions, aiming to improve the uncertainty estimates used by the recommendation process. Together, this thesis advances uncertainty-aware recommendation by showing how uncertain scores can be modeled and used to generate, adapt, and evaluate high-quality top-K recommendations.